Lured by the Consensus -...
Transcript of Lured by the Consensus -...
Lured by the Consensus
Roni Michaely, Amir Rubin, Dan Segal, and Alexander Vedrashko*
Abstract: We find that investors are fixated on analysts’ consensus outputs (earnings forecasts, recommendations,
and forecast dispersion), which can be inferior signals compared to the corresponding outputs provided by
high-quality analysts, especially when a large number of high-quality analysts follow the firm. This result,
which holds at the firm and market level, implies inefficient use of the information contained in analysts’
outputs. Further, the post-earnings announcement drift (PEAD) phenomenon occurs only when high-quality
analysts are more uncertain about the firm’s performance than all analysts following the firm. We conclude
that the market’s fixation on consensus measures has significant negative economic implications.
Keywords: consensus, analyst quality, forecasts, post-earnings announcement drift, stock
recommendations
JEL: G10; G11; G14; G17; G24; M41
*Michaely is from The Geneva Finance Research Institute, University of Geneva, ([email protected]). Rubin
is from Simon Fraser University and IDC ([email protected]). Segal is from IDC and Singapore Management University
([email protected]). Vedrashko is from Simon Fraser University ([email protected]). We thank the discussants and seminar
participants at the AsianFA meeting in Tokio, EFMA meeting in Milan, FIRS meeting in Barcelona, MFS meeting in
Budapest, Hebrew University, Tel-Aviv University, University of Connecticut, and University of Vienna for helpful
comments and suggestions.
1
Analysts are recognized as important information providers in financial markets. Investors
and academics alike use analysts’ consensus earnings forecasts as a measure of the market
expectations of firms’ future earnings. The perceived importance of the consensus forecast has
increased in recent years to the extent that even companies’ investor relations departments follow
the consensus on a continuous basis (Consensus earnings estimates report, 2013). Consensus
analysts’ recommendations have also become readily and, often, freely available for investors,
e.g., on the Nasdaq website. However, by construction, the consensus ignores the possibility that
analysts have different abilities and information sets, resulting in varying levels of forecast
accuracy and recommendation quality that are persistent over time.1
Thus, fixation on the consensus can have negative economic implications: investors may
be missing information and incur opportunity costs if they focus on the consensus and disregard
the heterogeneity in analyst quality. On the other hand, using the single number of the consensus
instead of differentiating among analysts can be optimal for investors. A simple average construct
can achieve a better level of predictability than more complex models that incorporate quality
differences (e.g., a literature review in Clemen, 1989); for example, the highest quality analyst’s
forecast can be less accurate than the consensus forecast, which diversifies individual errors.2 Even
if some analysts are consistently more accurate than the consensus, the costs of gathering
information and spending cognitive effort on identifying analysts with superior ability may exceed
1 Individual analysts’ forecast accuracy systematically differs for reasons including analysts’ varying experience
(Mikhail, Walther, and Willis, 1997; Clement, 1999; Hirst, Hopkins, and Wahlen, 2004), aptitude (Jacob, Lys and
Neale, 1999), education (Maines, McDaniel, and Harris, 1997; De Franco and Zhou, 2009), brokerage house
association and underwriting relationships (Lin and McNichols, 1998; Clement, 1999), proximity to the firm (Malloy,
2005), lead analyst and star status categorization (Stickel, 1992; Cooper, Day, and Lewis, 2001), or work habits
(Rubin, Segal, and Segal, 2017). 2 For example, consider a firm followed by four analysts, whose rankings in terms of forecast accuracy are perfectly
persistent over time. Let the top analyst’s forecast error (the analyst’s forecast minus actual earnings per share) be 2¢,
and the remaining three analysts have forecast errors of -4¢, -5¢, and 8¢. The consensus forecast error is 0.25¢, less
than the forecast error of the most accurate analyst and that of the average forecast of analysts in the top half of the
ranking, who can be called high-quality analysts. Since the persistence of analysts’ forecast accuracy is less than
perfect in actual markets, relying on the consensus forecast can make even more sense for investors.
2
the economic benefits of going beyond the consensus, such as when analyst quality heterogeneity
is small. In fact, investors’ fixation on the mean of analysts’ forecast distribution can be an instance
of central fixation bias, which describes people’s natural tendency to fixate their vision at the center
of a group of objects and which can be optimal for initial information processing (Tatler, 2007).
Hence, our objectives are two-fold. We first assess whether investors mainly focus on the
consensus output. Second, we examine whether analysts’ consensus outputs (e.g., consensus
earnings forecasts and consensus recommendations) are the right focal points for investors.
Building on the concept of diversification borrowed from asset pricing, we take into account not
only heterogeneity and persistence in analyst performance, which motivate the literature
contrasting high-quality (HQ; e.g., more accurate, “star”, “All-American”) and low-quality (LQ)
analysts, but also the diversification benefits of relying on the consensus, which can be a first-
order effect.3 Consequently, we advance the literature by considering the market’s focus on
consensus-type information and addressing the question of market pricing and economic efficiency
of consensus outputs. Further, while each of the prior studies examines only one of the outputs of
analysts and only at the firm level, we address our research question based on several outputs—
earnings forecasts, recommendations, and forecast dispersion—at both the individual firm and
market levels, which allows us to reach a general conclusion regarding the phenomenon of
investors’ consensus fixation.
Overall, we find that investors generally focus on the consensus, which leads to price
inefficiency. Differentiating among analysts and following HQ analysts instead of the consensus
can provide economic benefits. Our main results can be summarized as follows: first, the market
3 The literature has found inefficiencies in how the market treats analyst heterogeneity in reacting to earnings forecast
revisions (Stickel, 1992; Clement and Tse, 2003; Gleason and Lee, 2003; Chen, Francis, and Jiang, 2005) and
recommendations (Michaely and Womack, 1999; Mikhail, Walther, and Willis, 2004; Loh and Mian, 2006; Sorescu
and Subrahmanyam, 2006; Ertimur, Sunder, and Sunder, 2007; Fang and Yasuda, 2014; Kucheev, Ruiz, and
Sorensson, 2017).
3
reacts to the consensus forecast rather than to the more accurate average forecast generated by HQ
analysts, a categorization based on their past performance. This inefficient use of information in
analysts’ forecasts results in stock mispricing around earnings announcements. Second, we find
that it is not optimal for investors to follow the consensus recommendation revision compared to
the average recommendation revision of HQ analysts. Third, the dispersion of HQ analysts’
forecasts is more informative about the firm’s underlying uncertainty than the dispersion based on
all analysts following the firm (“the consensus dispersion”). Fourth, we find that the post-earnings
announcement drift (PEAD) arises only when the dispersion of HQ analysts is high relative to the
consensus dispersion; thus, investors who do not differentiate among analysts forego significant
returns. Finally, we find that HQ analysts’ recommendation revisions and forecast dispersion
aggregated over all firms in the sample predict the next month’s stock market return and volatility,
respectively, while the corresponding relations are not observed for the consensus
recommendations and consensus forecast dispersion. Taken together, our findings on earnings
announcements, recommendations, forecast dispersion, and the PEAD share an underlying
economic mechanism, which we refer to as the consensus fixation phenomenon, that causes
investors to systematically underweight quality differences among analysts, leading to pricing
inefficiency, and to forego investment returns as a result.
Following the literature (e.g., Chen, Francis, and Jiang, 2005; Loh and Mian, 2006), we
use the firm-year ranking approach and define HQ and LQ analysts according to their forecast
accuracy rankings for the previous year’s annual earnings announcement by the firm. We find
persistence in forecasting ability across time and firms an analyst covers, indicating that
forecasting performance captures analysts’ quality. Next, consistent with the diversification
effect—the more HQ analysts, the smaller the diversification advantage of the consensus relative
to the HQ analysts’ average forecast—we find that the average of HQ analysts’ forecasts is more
accurate than the consensus forecast only when a sufficient number of HQ analysts follow the firm.
4
Despite the superiority of the average of HQ analysts’ forecasts for these announcements, the
market reacts more vigorously to earnings surprises that are measured based on the consensus
forecast than the average of HQ analysts’ forecasts. This finding implies that the attention the
market pays to the consensus forecast may be excessive, and investors fail to fully incorporate the
information available in HQ analysts’ forecasts. The market’s fixation on the consensus forecast
is inefficient and can be exploited. A strategy based on the difference between the mean forecast
of HQ analysts and the consensus prior to the earnings announcement day yields economically and
statistically significant abnormal returns over the announcement day and the following trading day.
This phenomenon manifests itself in other aspects of analysts’ informational output, such
as recommendation revisions and forecast dispersion. We find that the market does not efficiently
impound HQ analysts’ recommendation revisions into prices. HQ analysts’ recommendation
revisions are strong predictors of the stock returns in the following month, in contrast to the
consensus recommendation revisions. This results in an opportunity cost for investors who do not
take into account analyst quality heterogeneity and who use investment strategies based on the
consensus recommendation, with a straightforward long-short strategy yielding a 1.1% monthly
alpha. Concerning analysts’ forecast dispersion, if the average of HQ analysts’ forecasts is more
informative than the consensus forecast, then the variability of their forecasts should also be a
superior measure of uncertainty regarding future firm performance relative to the consensus
dispersion. Indeed, we find that the dispersion of HQ analysts’ forecasts is a strong predictor of
the firm’s stock return volatility in the month following the annual earnings announcement month,
unlike the dispersion based on all analysts.
Building on studies showing that the PEAD arises in periods of uncertainty (Abarbanell,
Lanen, and Verrecchia, 1995; Mendenhall, 2004; Zhang, 2006a; Francis et al., 2007; Hung, Li,
and Wang, 2014) and using our finding that HQ analysts’ forecast dispersion can proxy for the
firm’s uncertainty better than the consensus dispersion, we find a significantly greater PEAD
5
(9.4% after 11 months) when HQ analysts are relatively uncertain than the PEAD in the full sample
of earnings announcements, which ignores analyst quality heterogeneity. Overall, these findings
indicate that the consensus fixation bias is relevant to both short (e.g., earnings announcements)
and long (the PEAD) investment horizon decisions. Investors recognizing the information
contained in HQ analysts’ forecasts are able to make significantly better investment decisions than
those investors making their investment decisions based on the consensus forecast.
Finally, we examine whether the inferiority of consensus outputs is relevant at the market
level. Since HQ analysts’ recommendation changes are better at the firm level, HQ analysts’
average recommendation change across all firms in the sample can also be a superior predictor of
the market return. We find that HQ analysts’ average recommendation changes predict the market
returns for the following month, in contrast to the consensus recommendation changes. For
example, a long-short strategy based on the direction of HQ analysts’ recommendation revision
relative to the consensus revision produces a 6.8% annualized calendar-time alpha in the month
following the earnings announcement. The argument for the average dispersion of HQ analysts’
forecasts rather than the consensus dispersion predicting market volatility is analogous, and we
find that HQ analysts’ normalized forecast dispersion predicts the next-month CBOE volatility
index (VIX) and changes in the VIX, thereby allowing investors unimpeded by consensus fixation
to earn higher returns.4
Our findings contribute to the literature across several dimensions. First, we analyze the
market efficiency implications of analyst quality differences not previously considered in the
literature. These implications include the short- and long-term reaction to earnings announcements,
4 We also note that aggregating analysts’ informational outputs about firms they follow is akin to surveys aggregating
the opinions of representative households to measure consumer confidence (Ludvigson, 2004), both of which can be
used to predict macroeconomic and market-level variables.
6
generating a more informative forecast dispersion measure, and finding the predictability of market
returns and volatility based on the aggregation of analysts’ firm-level outputs.
Second, the study suggests that investors’ focus on the consensus earnings forecast may be
a result of limited attention (e.g., Hirshleifer, Lim, and Teoh, 2009). Investors may prefer the
expediency of the single number of the consensus to the exertion of cognitive effort to assess
analyst quality. From this perspective, our finding that investors ignore the valuable information
contained in HQ analysts’ outputs contributes to the literature that uses subjective expectations of
bond risk premia (Buraschi, Piatti, and Whelan, 2017), corrections for biases in individual
analysts’ forecasts (Chiang et al., 2018), and firm fundamentals (Wahlen and Wieland, 2011; So,
2013) to show that investors can do better than relying on consensus expectations.
The final aspect of our contribution applies to several strands of literature. First, our insight
that different analysts bring different amounts of information to the market advances the debate on
whether recommendations are informative (Barber et al., 2001; Jegadeesh et al., 2004; Altinkilic
and Hansen, 2009; Loh and Stulz, 2011). Second, our findings limit the extent of the PEAD
anomaly’s challenge to market efficiency (Fama, 1998), in that the PEAD is restricted to periods
of high uncertainty regarding the firm’s prospects. Third, our findings on the PEAD and aggregate
volatility suggest that high- and low-uncertainty firms should be identified based on the forecast
dispersion adjusted for analyst quality differences rather than using the consensus-based dispersion
alone (e.g., Diether, Malloy, and Scherbina, 2002; Johnson, 2004; Barron, Stanford, and Yu,
2009).
2. Data and variables
We use the sample of annual earnings per share (EPS) estimates and earnings
announcements in I/B/E/S during January 1992-December 2015 for companies with daily return
7
data provided by the Center for Research in Security Prices (CRSP) database.5 The starting year
of 1992 is chosen because some analyses require analysts’ recommendation data, which begin in
1993. The earnings estimates and actual earnings are adjusted for splits using the daily cumulative
adjustment factor from CRSP (Glushkov and Robinson, 2006).
Each year, we rank analysts following the firm based on the closest absolute forecast error,
which is the absolute difference between an analyst’s earnings forecast closest to the earnings
announcement (made at least a day prior to the announcement day) and the actual earnings, divided
by the share price at the beginning of the calendar year. We define HQ (LQ) analysts based on
whether their absolute closest forecast error for the firm-year is below (above) the median absolute
forecast error for the firm-year. The firm-year ranking measure is directly suited for our study
compared to various “star” analyst, cross-firm classifications (e.g., the Wall Street Journal’s “Best
on the Street” or the Thomson Reuters StarMine’s “Top Earnings Estimators”) in that it has the
cross-firm persistence property similar to the star analyst classifications, as we find below, but it
is also simpler (uses only analysts’ prior accuracy in a given firm) and preserves the sample size
because many firms are not followed by such star analysts. In the robustness section, we conduct
sensitivity tests, whose results indicate that our findings are not affected by different cutoff values
or forecast weights in the definitions of HQ and LQ analysts. It is important to note that advancing
the literatures on measuring analysts’ quality, determining the subset of superior analysts, or
analyzing their relative performance is not among our study’s objectives. Instead, we group
analysts by their quality with as simple, realistic, and robust alternative to the consensus, and our
contribution is to examine the economic implications of investors’ ignoring variation in analyst
quality and using the consensus data instead.
5 We focus on annual rather than quarterly earnings for two main reasons. First, fewer analysts provide quarterly
forecasts than annual forecasts. Second, annual earnings announcements are typically more informative, including
that they are more often supplemented with a conference call and followed by recommendation changes.
8
From the initial sample, we generate 861,349 firm-year-analyst rankings based on the
closest forecast error; the number drops to 804,003 observations once we require firms to have
Compustat data. Next, to avoid small sample bias in our ranking when the number of analysts
following the firm is small, we exclude firm-years with fewer than four analysts following, thereby
reducing the sample to 750,295 observations. In addition, for all analyses except those using stock
recommendations (Tables 4 and 8), analysts must appear in the data in two consecutive years for
a given annual announcement to be ranked based on the first of the two years, reducing the sample
to 485,815 observations.
In the firm-level regressions, we control for the following firm characteristics: size, the
annual stock return, the book-to-market ratio, the number of analysts following, and leverage. Firm
size is the market value of the firm’s equity at the end of the month prior to the earnings
announcement month. The annual stock return is measured based on monthly equity returns in the
12 months prior to the earnings announcement month. The book-to-market ratio is computed as
stockholder equity minus preferred stock plus deferred taxes at the end of the fiscal year for which
the earnings are announced and divided by firm size. The number of analysts is the number of
analysts who made at least one earnings forecast for the given announcement. Leverage is the book
value of total liabilities divided by total assets at the end of the fiscal year for which the earnings
are announced. Some of the regression models also control for analyst characteristics: (i) overall
tenure is the number of years since the analyst first appeared in the I/B/E/S file; (ii) firm-specific
tenure is the number of years since the analyst began covering the company in the I/B/E/S file;
(iii) brokerage house size is the number of analysts employed by the brokerage firm; and (iv) firm
coverage is the number of firms covered by the analyst.
In the models predicting market returns and volatility, most of the controls we use follow
Li, Ng, and Swaminathan (2013) and are measured in the month prior to the dependent variable’s
month. The earnings-to-price ratio and dividend-to-price ratio are calculated from the S&P 500
9
dividend, earnings, and price data on Robert Shiller’s website.6 The one-month T-bill rate and 30-
year Treasury yield are obtained from CRSP. The term spread is the difference between the AAA-
rated corporate bond yields obtained from the Federal Reserve Bank of St. Louis (FRED) database
and the one-month T-bill yield. The default spread is the difference between the BAA and AAA
corporate bond yields for the last day of the month when both BAA and AAA daily yields exist,
obtained from the FRED. Inflation is the change in the consumer price index (CPI; all urban
consumers, monthly, non-seasonally adjusted) obtained from the FRED. Following Da, Engelberg,
and Gao (2015), our regressions of the VIX also control for the perceived economic policy
uncertainty (EPU), which is a news-based measure provided by Baker, Bloom, and Davis (2016).
EPU change is the percentage change in the monthly average of the daily EPU for the month prior
to the dependent variable’s month. The VIX is from Wharton Research Data Services (WRDS).
3. Individual analysts
3.1. Sample description and tests of persistence in forecasting performance
Our preliminary analysis determines whether HQ and LQ analysts’ forecast accuracy is
persistent, which is the prerequisite for using groups of analysts ranked according to our quality
measure as alternatives to the consensus. The unit of observation in all analyses in this section is
the individual analyst. Figure 1 shows the mean absolute forecast errors of HQ and all analysts for
each day during the 300 days prior to the earnings announcement. We observe acceleration in the
reduction of the mean forecast error around quarterly earnings announcements at the -90, -180,
and -270 day marks. The graph shows that the mean absolute forecast error of all analysts is higher
than the mean absolute forecast error of HQ analysts in all days prior to the earnings
announcement. This result indicates that the analyst ranking is persistent in the time dimension
6 Available at http://www.econ.yale.edu/~shiller/data.htm.
10
because the ranking is based on the previous year’s accuracy. The mean absolute forecast errors
of All and HQ analysts decrease over time to approximately 0.012 and 0.0115, respectively, one
day before the earnings announcement. This difference of 4.17% (0.0115
0.0120− 1) is economically
meaningful and statistically significant (p-value<0.01). Notably, HQ analysts’ accuracy 30 days
before the announcement is already higher than all analysts’ accuracy at the announcement.
Table 1 reports the summary statistics for analysts ranked into the LQ and HQ groups and
examines the persistence of their forecast accuracy. In Panel A, we find that relative to LQ analysts,
HQ analysts tend to be more experienced overall and within individual firms, be employed by
larger brokerage firms, and cover a greater number of firms. To analyze the univariate persistence
of analysts’ forecast accuracy, we compare HQ and LQ analysts’ forecast errors at the earnings
announcement, i.e., a year after they were ranked. The absolute forecast errors of HQ analysts
remain smaller than those of LQ analysts: the difference is 9% (0.0081/0.0089) and statistically
significant. In the last line of Panel A, we find that both HQ and LQ analysts’ forecasts exhibit
approximately equal magnitudes of optimism bias; the average forecast errors are significantly
different from zero, with untabulated p-values<0.01, indicating that HQ analysts’ greater forecast
accuracy does not appear to be associated with a more positive forecast bias (Lim, 2001).
The regression analysis in Panel B of Table 1 examines the persistence in the quality
classification of analysts over time (columns (1)-(4)) and across firms (columns (5) and (6)). In
the probit models in columns (1) and (2), the dependent variable is the HQ status in year t, which
equals one if the analyst is categorized as HQ and zero otherwise. In columns (3) and (4), the
dependent variable is the absolute forecast error, a continuous variable that allows us to control for
firm fixed effects in the regression. In columns (1) and (3), we control for firm characteristics, and
in columns (2) and (4), we control for both firm and analyst characteristics. The main coefficient
of interest is the HQ classification in year t-1. The results show that the coefficient on HQ status
11
(t-1) is highly significant (p-value<0.01) in all specifications, indicating that analysts’ rankings
and forecast accuracy are persistent in consecutive years. For example, the unconditional
probability of belonging to the HQ group is approximately 50%, and according to columns (1) and
(2), accounting for the HQ status in the previous year increases this likelihood by approximately
4.1%. Columns (3) and (4) show that HQ analysts continue to have lower absolute forecast errors
in the following year, with their average absolute forecast error being 8.2% lower (0.00072/0.0085)
than the average absolute forecast error for all analysts.
We next conduct cross-firm tests to examine whether forecasting performance is persistent
not only over time but also across firms that the analysts follow. Not only is this analysis important
in its own right, but affirmative findings will also reinforce the argument that some analysts
provide superior information than do others; that is, the HQ designation is not a firm-specific
attribute but, rather, a characteristic of the analyst. We define an analyst’s performance in other
firms as HQ if the analyst is classified in the HQ category for the majority of the other firms that
he or she follows during the year (excluding this firm).7 Columns (5) and (6) of Panel B test
whether ranking as an HQ analyst in the other firms in year t-1 can predict an analyst’s forecasting
performance in year t over and above the HQ classification in year t-1 in the same firm. We
estimate two probit models where the dependent variable is the HQ status indicator in a given firm
in year t. The independent variables of interest are the HQ status indicator of this analyst in the
same firm in year t-1 and the HQ status in other firms indicator, which is equal to one if this analyst
is also HQ in the majority of other firms that he or she followed in year t-1. We find that analysts
who were of HQ in the majority of other firms they followed in year t-1 are 5.1% (p-value<0.01)
more likely to be HQ in a given firm in year t. The coefficient on the firm-specific HQ designation
7 If the number of HQ and LQ rankings of the analyst in the other firms is the same, this analyst-year-firm observation
cannot be categorized as either HQ or LQ in the other firms and is thus excluded from this analysis (approximately
9% of the observations).
12
in year t-1 remains positive and significant (p-value<0.01). Hence, the cross-firm findings suggest
that analysts’ forecasting performance transcends across the stocks that they follow and, further,
that HQ analysts are indeed better than their peers in a persistent manner.
3.2. What does persistence in individual forecasting performance mean for the consensus?
Our finding that individual HQ analysts tend to persistently provide more accurate earnings
forecasts than LQ analysts seems to imply that the consensus may be an inferior predictor of future
earnings than the average of HQ analysts’ forecasts. However, the actual extent of how much the
average of HQ analysts’ forecasts is more accurate than the consensus forecast in a specific firm
may depend on the number of HQ analysts following the firm.8 Table 2 empirically investigates
this issue in Panel A and provides statistical tests comparing the absolute standardized unexpected
earnings of consensus (SUE of consensus, equal to the difference between the actual earnings and
the average forecast provided by all analysts following the firm, normalized by the stock price)
with the absolute SUE of HQ analysts, which is based on the difference between the actual earnings
and the HQ analysts’ average forecast.9 We find that as the number of HQ analysts following a
firm increases, HQ analysts as a group eventually become more accurate than the consensus,
confirming our conjecture. Further, when the number of HQ analysts is four or more, the absolute
forecast error of HQ analysts is smaller than the consensus. Therefore, it is in these firms that
8 The intuition is simple—in averaging a greater number of forecasts, the noise portion of their individual forecasts
cancels out more, leading to a more precise forecast of the true earnings signal. The greater the number of HQ
analysts following the firm is, ceteris paribus, the smaller the diversification advantage of the consensus is and more
likely investors are to obtain a more accurate forecast by following the average forecast of the HQ analysts than they
would via the consensus. 9 Because some HQ and LQ analysts may stop covering the firm after year t-1 and new, unranked analysts may
commence coverage, the numbers of HQ and LQ analysts in year t can become too small or too different from each
other (e.g., five HQ and one LQ or vice versa), leading to small sample bias and a lack of robustness when the average
accuracies of HQ, LQ, and all analysts as groups are compared in the firm-level analysis. To mitigate this issue, we
restrict the sample in all firm-level analyses (Tables 2-7) to firm-years in which the numbers of HQ and LQ analysts
are not too different in year t. Specifically, we require that neither of these groups exceeds 75% of all analysts
providing forecasts for a given announcement.
13
investors seeking more accurate earnings forecasts should forego the consensus forecast in favor
of the average of HQ analysts’ forecasts. For the same reason, our sample consists of firms with
four or more HQ analysts when we examine whether the market can exploit differential analyst
quality in the analysis of recommendation changes, forecast dispersion, and the PEAD.
In Panel B of Table 2, we report firm characteristics for the samples with fewer than four
and four or more HQ analysts in year t, corresponding to firm-years for which the consensus
forecast is superior or inferior to HQ analysts’ average forecast, respectively. Not unexpectedly,
firms with four or more HQ analysts are significantly larger. While these firms represent less than
50% of all firms, they account for 87% of the market value of the full sample. Firms with four or
more HQ analysts are more levered and perceived by the market to have a greater growth potential
(according to the book-to-market ratio of 0.66, compared to 1.36 for firms with few HQ analysts).
This association between firm characteristics and the number of analysts following (or the number
of HQ analysts) is not surprising and consistent with prior literature. It is important, however, in
the context of our findings, suggesting that the sample for testing the inefficiency of consensus
fixation is not dominated by small firms or firms followed by few or inexperienced analysts.
4. Earnings announcements
The previous section demonstrates that relying on HQ analysts’ earnings forecasts can
generate an earnings forecast that is superior to the consensus forecast. To test whether the market
is aware of this empirical regularity, we examine whether the immediate reaction to the earnings
surprise based on the mean forecast of HQ analysts is greater than the reaction to the earnings
surprise based on the consensus forecast. Table 3 reports the regression results in which the
dependent variable is the buy-and-hold cumulative abnormal return (BHAR) for the earnings
announcement day and the following trading day based on the four-factor model (Fama and
French, 1993; Carhart, 1997). The main variables of interest are the coefficients on the SUE based
14
on the consensus and HQ analysts. Table 3 shows that the reaction to the SUE based on the
consensus forecast is greater than the reaction to HQ analysts’ SUE, with a highly statistically
significant difference between the coefficients of 0.103 based on the chi-squared test in the full
sample and a slightly smaller but still highly significant difference of 0.06 in the sample of firms
with four or more HQ analysts. The coefficient on HQ analysts’ SUE is greater and significantly
different from the coefficient on LQ analysts’ SUE, which suggests that the market is partially
aware of the accuracy differences among analysts.10 This finding is in line with Kirk,
Reppenhagen, and Tucker (2014), who show that the market reacts more strongly to the key analyst
than to the least influential analyst following the firm. Importantly, the results indicate that the
market does not sufficiently recognize analyst quality differences because its reaction to the
consensus forecast is significantly stronger even in columns (4)-(6), where HQ analysts are, on
average, more accurate than the consensus.
The finding that the market does not give sufficient weight to HQ analysts’ forecasts
relative to the consensus may have meaningful economic implications. To gauge their magnitude,
we first construct a simple measure of earnings surprise based on the difference between HQ
analysts’ mean forecast and the consensus forecast, labeled predicted surprise. The intuition is to
replace the actual earnings in the SUE formula with HQ analysts’ mean forecast,
𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑 𝑠𝑢𝑟𝑝𝑟𝑖𝑠𝑒 =𝐴𝑣𝑔.𝐹𝑜𝑟𝑒𝑐𝑎𝑠𝑡𝐻𝑄−𝐴𝑣𝑔.𝐹𝑜𝑟𝑒𝑐𝑎𝑠𝑡𝑐𝑜𝑛𝑠𝑒𝑛𝑠𝑢𝑠
𝑃𝑟𝑖𝑐𝑒𝑡−1, (1)
so that predicted surprise can be used to predict the SUE of consensus. Investors aware of the
quality differences among analysts would be able to use this measure to predict the immediate
market reaction to earnings announcements. Given that HQ analysts are more accurate than the
consensus and that the market overweights the consensus forecast when it reacts to earnings
10 As a technical note, because the set of analysts covering a firm tends to change after the ranking in the previous
year, the proportion of HQ and LQ analysts varies across firms and years. Therefore, SUE of consensus is not an
average of SUE of HQ and SUE of LQ analysts.
15
surprise, one can expect positive or negative abnormal returns to the earnings announcement when
the mean forecast of HQ analysts is greater or smaller than the consensus, respectively. A simple
hedging strategy assessing the predictability of the immediate reaction to earnings announcements
is to buy the stock at the market close on the day before the announcement when the predicted
surprise is positive and to short it when it is negative. Because the market is focused on the
consensus, it will be surprised the most when HQ analysts’ average forecast is different the most
from the consensus. Next, since the consensus includes new analysts, whose quality is currently
unknown because they started to follow the firm after the last year’s ranking, the difference
between HQ and LQ analysts’ forecasts measures analyst heterogeneity not priced because of
consensus fixation. Therefore, we also consider a strategy using an alternative definition for
predicted surprise based on the normalized difference between HQ and LQ analysts’ mean
forecasts:
𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑 𝑠𝑢𝑟𝑝𝑟𝑖𝑠𝑒 =𝐴𝑣𝑔.𝐹𝑜𝑟𝑒𝑐𝑎𝑠𝑡𝐻𝑄−𝐴𝑣𝑔.𝐹𝑜𝑟𝑒𝑐𝑎𝑠𝑡𝐿𝑄
𝑃𝑟𝑖𝑐𝑒𝑡−1 (2)
The potential returns that investors can earn using these alternative predicted surprise measures
represent the returns foregone for investors if they use the surprise based on the difference between
the consensus and actual earnings, thereby not incorporating analysts’ quality differences into
prices.
We report the empirical results in Table 4. The analysis is based on two variations of the
signal based on predicted surprise: the positive predicted surprise and big positive predicted
surprise indicators. Positive predicted surprise is equal to one if predicted surprise is positive and
zero otherwise. A stronger signal, the big positive predicted surprise indicator is one (zero) if
predicted surprise is above (below) the median of its positive (negative) values in the previous
year, and is set to missing otherwise. Using the values of predicted surprise measured in the
previous year ensures that our analysis is out-of-sample. We regress the two-day cumulative
16
BHAR on each of these indicators and control variables. The coefficients on the predicted surprise
indicators are positive and significant in all specifications, reaching 0.0019 in column (3), and the
statistical significance of the predicted surprise indicators is greater for the definition based on the
difference between HQ and LQ analysts’ forecasts. The last line of the table reports the two-day
abnormal returns of a hedging strategy that is long if the predicted surprise indicator of that column
is equal to 1 and short if it is equal to 0. All returns are statistically significant and reach 0.24%
for big positive predicted surprise based on the difference between HQ and LQ analysts’ forecasts.
These returns can be high enough relative to the transaction costs (Novy-Marx and Velikov, 2016)
because predicted surprise achieves its highest values when HQ analysts are most accurate, i.e.,
in firms followed by many analysts (according to Table 2), implying relatively small transaction
costs for these larger firms.
The overall conclusion from Tables 3 and 4 is that the market seems to overreact to the
actual earnings’ deviations from the consensus, compared to deviations from HQ analysts’ average
estimate. This fixation on the consensus forecast presents an opportunity cost for investors and
prevents stock prices from efficiently reflecting the available information.
5. Stock recommendations, forecast dispersion, and implications for the PEAD
The persistence in analysts’ forecasting performance over time and across stocks suggests
that HQ analysts have superior ability; thus, it is possible that they issue superior stock
recommendations compared to the consensus recommendation based on all recommendations for
the firm.11 Further, given that HQ analysts are better at forecasting future earnings, the dispersion
in their forecasts is likely to contain more relevant information than the dispersion of the forecasts
11 A relation between earnings forecasts and recommendations has been examined in the prior literature (e.g., Francis
and Soffer, 1997; Bradshaw, 2004; Loh and Mian, 2006; Ertimur, Sunder, and Sunder, 2007).
17
of the entire set of analysts following the firm. In this section, we empirically examine these
predictions.
5.1. Stock recommendations
To gauge whether investors can benefit from being aware of the differences in analysts’
forecasting ability, we examine how future returns are associated with the recommendation
revisions of different subsets of analysts. If investors internalize and act on the quality differences
among analysts, which would imply that they are not fixated on the consensus recommendation,
then no relation between future returns and recommendation revisions should obtain regardless of
the quality of the analysts making the recommendations.
A recommendation is an integer between 1 and 5, where 1 is “strong buy”, 5 is “strong
sell”, and 3 is “hold”. For ease of interpretation, we measure recommendation revisions as the
negative of the current recommendation of the analyst minus the previous recommendation of the
analyst; thus, a positive recommendation revision is an upgrade. The recommendation revision for
the firm is the average of individual analysts’ revisions. The sample consists of the
recommendation revisions made during the month of the annual earnings announcement. The
earnings announcement month has several advantages making it the best time frame to examine
whether the market efficiently incorporates its knowledge of analyst quality into its reaction to
recommendations. First, the month with the annual announcement has the most information for
analysts to process during the year because the information in earnings announcements has a major
influence on recommendation revisions (Yezegel, 2015). Second, analysts of different quality
types face the same information set when making recommendations that month, in contrast to
recommendations at random dates during the year. Finally, and perhaps most importantly, it is the
earnings announcement month that reveals that the market is fixated on the consensus forecast and
18
does not recognize superior HQ analysts; thus, we expect this pattern to be prominent for HQ
analysts’ stock recommendations during this month, as well.
We first examine the immediate market reaction to recommendation revisions, which is
inherently an individual analyst-level rather than analyst group-level event, i.e., there is no
consensus recommendation to consider in the analysis of the immediate reaction. The untabulated
regressions of the immediate market reaction to HQ and LQ analysts’ individual recommendation
revisions yield results that are consistent with the finding for earnings announcements: investors
recognize, at least to an extent, the more accurate forecasters by reacting more strongly to the
recommendation revisions of HQ analysts relative to those of LQ analysts. However, the important
question remains whether the market fully incorporates quality differences into prices or is fixated
on the consensus recommendation.
Table 5 reports the results concerning a delayed response to recommendation revisions by
different analyst types. For each calendar month with an annual earnings announcement by a given
firm, we average all, HQ, and LQ analysts’ recommendation revisions and regress the next months’
stock return on these recommendation revision averages. This analysis predicting equity returns
one calendar month ahead enables us to use all revisions during the current month because, by the
end of that month, investors have learned and updated the analyst quality classification based on
this new earnings announcement.12 The investment delay from the revision date to the end of the
revision month provides investors with sufficient time to react to the revision and, because such a
12 The updated analyst ranking also allows all the revisions to be used because all analysts following the firm are
ranked as of the end of the announcement month, i.e., there are no revisions by unranked, newcomer analysts who
have just started following the firm between the earnings announcement and the end of the month. The sample has
exactly 50% HQ and 50% LQ analysts; thus, the requirement for earnings announcements that at most 75% of forecasts
are made by one analyst type is not binding for recommendation revision analyses.
19
delay reduces the market response the next month, leads to conservative return estimates for the
next month (Barber et al., 2001).13
The regression results in columns (1)-(3) of Table 5 reveal that the coefficients on the
consensus and HQ recommendation revisions are positive and significant. The greater coefficient
for HQ analysts and not significant coefficient for LQ analysts imply that price momentum
following consensus recommendation revisions (e.g., Jegadeesh et al., 2004) is driven by HQ
analysts’ recommendation revisions.14 This supports the argument in the literature that analyst
recommendations bring new information to the market (Jegadeesh et al., 2004; Loh and Stulz,
2011).
To examine whether investors should follow the HQ recommendation revision rather than
the consensus recommendation revision, we regress the next month’s return on the difference
between HQ analysts’ and consensus recommendation revisions. The intuition for this test is that
the gain for investors is the greatest when HQ analysts are relatively more informed, i.e., when
their average recommendation is more different from the consensus recommendation. The
regression result in column (4) shows that following HQ analysts’ upgrades or downgrades net of
the consensus revision leads to higher out-of-sample returns for investors.
To further assess the additional value inherent in incorporating analyst quality differences
into investment decisions, we examine the returns of a long-minus-short strategy in the month
following the revisions, where the long (short) position is in the firms for which the consensus,
HQ, LQ, or HQ minus the consensus mean recommendation revision variable is positive (negative)
during the earnings announcement month. The last two lines of Table 5 report both the event-time
13 The results for predicted monthly returns in Table 5 are unaffected by using the subsamples of recommendation
revisions made either before or after the earnings announcement during that calendar month or coinciding with the
announcement’s two-day window. We also reach the same outcome by conducting event-time analysis using returns
over the periods (2,32) and (2,62) days following the revision. 14 Loh and Mian (2006) and Ertimur, Sunder, and Sunder (2007) contrast market responses to the recommendations
of more and less accurate analysts based on a partially contemporaneous rather than predictive relation between the
recommendations of the two quality groups and stock returns.
20
average long-short return and the four-factor alpha based on the calendar-time approach for this
strategy. In particular, with HQ analysts’ recommendation revisions, the resulting one-month
strategy return is 0.85% and highly statistically significant, in contrast to the consensus
recommendation revisions, for which the hedging strategy yields a statistically nonsignificant
0.36%. For the HQ net of the consensus trading signal, the return is even higher, at 1.24% per
month. The calendar-time approach finds that investors achieve a significant alpha of 1.06%
(12.72% annualized) when they rely on HQ analysts’ recommendation revision net of the
consensus recommendation revision.
Overall, the predictive relation between analyst recommendation revisions and equity
returns in the subsequent month is driven by HQ analysts’ mean recommendation rather than the
consensus recommendation. Hence, our findings suggest that analyst quality based on earnings
forecasts generalizes to recommendation revisions and that investors can earn significant
additional returns by taking into account analyst quality differences instead of relying on the
consensus recommendation. These conclusions are entirely consistent with the notion that treating
all analysts as equal and, thus, using the consensus measure of analyst output can lead to inefficient
pricing.
5.2. Analysts’ forecast dispersion and volatility
Analysts’ forecast dispersion has been widely used as a proxy for uncertainty about firms’
future prospects. We conjecture that just as HQ analysts’ superior earnings forecasts and
recommendations indicate that they have superior information concerning firm value, these
analysts’ forecast dispersion similarly contains more accurate information about future
uncertainty. We examine whether the disagreement about the firm’s prospects among HQ analysts
is a superior predictor of uncertainty surrounding the firm’s future performance, measured by
21
future return volatility, relative to the disagreement among all analysts, which we call the
consensus dispersion.
Table 6 reports the results of regressing stock return volatility during the month following
the earnings announcement month on the standard deviation of analysts’ forecasts before the
earnings announcement. To avoid stale forecasts and to make forecasts more comparable in terms
of their proximity to the announcement, we use only the forecasts in the 60 days prior to the
announcement.15 We consider the dispersion of forecasts by all, HQ, and LQ analysts separately,
whose coefficients are the variables of interest. HQ analysts’ forecast dispersion is statistically
significant, while the consensus dispersion is not significant. These findings suggest that only HQ
analysts’ forecasts capture variation in uncertainty, which is associated with future equity volatility
in a given firm. This result extends the earlier findings of a contemporaneous relation between
analysts’ forecast dispersion, stock volatility, and the option value of the firm (Ajinkya and Gift,
1985). The next two subsections illustrate how investors can benefit from utilizing HQ analysts’
dispersion instead of the consensus dispersion.
5.3. Post-earnings announcement drift
Our results concerning the differences in the properties of forecast dispersion of all and
HQ analysts’ forecasts have important implications for the PEAD anomaly. First, the model
proposed by Abarbanell, Lanen, and Verrecchia (1995) predicts that when forecast dispersion is
high, investors place less weight on the forecasts relative to their private information, so that
investors reduce their immediate response to earnings surprise. Consequently, following the
15 The length of the measurement period for forecast dispersion significantly varies in the literature. For instance, it
can be one month (Krishnaswami and Subramaniam, 1999), four months (Zhang, 2006b), six months (Babenko,
Tserlukevich, and Vedrashko, 2012), and up to one year since the previous earnings announcement (Diether, Malloy,
and Scherbina, 2002). Our choice of 60 days ensures that we use only the annual earnings forecasts made after the
previous quarterly earnings announcement. Our results are unaffected by using a different period length.
22
announcement, as investors receive more information over time, prices adjust in a manner
potentially resulting in a PEAD. Similarly, Zhang (2006a) argues that investor underreaction to
public information is more significant when uncertainty is high and finds that analysts’ forecast
dispersion predicts a price drift following analysts’ forecasts.16 Second, our finding of a relation
between HQ analysts’ dispersion and return volatility, combined with a relation between return
volatility and the PEAD documented in Mendenhall (2004), posit a link between HQ analysts’
dispersion and the PEAD. For the full set of analysts, this relation is expected to be weaker because
of a weaker relation between consensus dispersion and return volatility.
Therefore, we examine whether investors can better exploit the PEAD when the dispersion
of HQ analysts’ forecasts is high and relative to the consensus dispersion than when they disregard
analyst quality differences.17 We calculate the PEAD using the calendar-time approach. To be able
to relate our results to the PEAD literature, we use the consensus earnings surprise measure to
assign announcing stocks to a long or short portfolio at the end of each month, depending on
whether the earnings surprise (SUE of consensus) is positive or negative, respectively. The stocks
are then held in the two portfolios for horizons from 1 to 11 months to avoid overlapping with the
next annual earnings announcement. The monthly PEAD is the alpha from regressing the monthly
value-weighted portfolio returns on the four Fama-French and Carhart factors.18 The cumulative
PEAD is the monthly alpha for a given horizon (the number of months that the stock is held in the
16 In related studies, Francis et al. (2007) find a positive relation between uncertainty, which they measure with the
unexplained portion of working capital accruals, and the PEAD. Hung, Li, and Wang (2014) find that exogenously
reduced information uncertainty leads to a lower PEAD. 17 For an association between the PEAD and HQ analysts’ forecast dispersion to exist, investors do not need to be
aware or take into account HQ analysts’ forecast dispersion because they can learn about the firm’s uncertainty level,
including its systematic component, from a variety of sources. HQ analysts’ forecast dispersion effectively reflects
that information. 18 We obtain similar results using equal-weighted portfolios. Separately, we also note that because of our sample’s
requirement that four or more HQ analysts follow the firm, the sample consists of relatively large firms, thereby
reducing the likelihood that any findings may be attributed to illiquidity (Sadka, 2006).
23
long or short calendar-time portfolio) multiplied by the horizon length in months. The resulting
relation between forecast dispersion and the PEAD is presented in Figure 2 and Table 7.
Figure 2 reports the cumulative long PEAD portfolio return minus the short PEAD
portfolio return for the sample of announcements with high uncertainty, defined as announcements
for which HQ analysts’ forecast dispersion is greater than the consensus dispersion, the low-
uncertainty sample, in which HQ analysts’ forecast dispersion is lower than the consensus
dispersion, and the full-sample, i.e., regular, PEAD. The high-uncertainty PEAD is clearly greater
than the full-sample PEAD, and the low-uncertainty PEAD is below the full-sample PEAD. The
largest cumulative long-short abnormal return of the PEAD strategy is 11.2%, reached eight
months following high-uncertainty announcements. Table 7 reports the monthly alphas on long,
short, and long-minus-short strategies for the subsamples with high- and low-uncertainty
announcements. We find that the low-uncertainty PEAD alpha (approximately 60% of the
announcements) is not statistically significant except for the 11-month horizon. In contrast, when
the forecast dispersion of HQ analysts is greater than the consensus dispersion, the abnormal return
of the long-minus-short portfolio is highly statistically significant at almost all horizons. To
determine the additional return that investors can earn from this strategy over and above the return
from the simple PEAD strategy, we regress the long-short post-announcement raw return for the
high-uncertainty portfolio on the four Fama-French and Carhart factors and the long-short post-
announcement raw return for the portfolio including all announcements. The resulting monthly
alphas for different horizons are reported in column (7). Investors can earn up to 9.4% cumulative
alpha (at the eight-month horizon) over and above the regular PEAD alpha. Overall, the PEAD is
observed primarily during periods of high information uncertainty, as determined by the forecast
dispersions of HQ analysts, which allows investors who are not fixated on the consensus forecast
and recognize analyst quality differences to achieve better investment performance.
24
6. The content of HQ analysts’ information output at the aggregate
Our findings that HQ analysts’ have superior ability to forecast earnings, issue more
informative recommendations, and have better forecasts of volatility have immediate implications
for aggregate returns and volatility: we should expect that the superior abilities at the firm level
manifest themselves at the aggregate market levels, thereby benefiting investors who overcome
their consensus fixation. Accordingly, we average the changes in recommendations of all, HQ, and
LQ analysts across all firms that made an annual earnings announcement this month and examine
whether the consensus or HQ analysts’ average recommendation revisions predict future market
returns. The aggregation argument works similarly for forecast dispersion, in that aggregating HQ
analysts’ dispersion across all firms in the sample should result in a dispersion measure reflecting
the degree of uncertainty in the market.19
In Table 8, Panel A, we report the estimation results for the relation between revisions in
stock recommendations and future market returns. To that end, each month, we calculate a simple
average of HQ, LQ, and all analysts’ recommendation changes across all firms that announce
annual earnings, resulting in one monthly time series for the average recommendation change in
the market. The dependent variable is the value-weighted market return in the month following the
recommendation change month. The consensus mean recommendation revision is the mean of all
recommendation changes during the month in which the firm's earnings are announced. The HQ
(LQ) mean recommendation revision is an analogous variable that is based only on
recommendation changes of HQ (LQ) analysts. The HQ-consensus mean recommendation
revision is equal to the difference between the HQ analysts’ and consensus mean recommendation
revisions when the HQ analysts’ mean revision is either positive and greater or negative and
19 Notably, the aggregation principle does not assume and the aggregate findings do not indicate that individual HQ
analysts have superior macroeconomic knowledge or ability to predict market-level developments (e.g., Hutton, Lee,
and Shu, 2012).
25
smaller than the consensus revision; otherwise, this variable is zero. The control variables follow
Li, Ng, and Swaminathan (2013) and also include the previous month’s market return to account
for return momentum.
The regression results reveal that the coefficient on the consensus recommendation revision
is positive and weakly statistically significant, indicating that the information in recommendation
revisions is not fully internalized by the market. However, this coefficient’s significance is entirely
due to the mean recommendation revision by all HQ analysts because LQ analysts’ coefficient is
not statistically significant. The coefficients on the HQ revision net of the consensus revision in
columns (4) and (8) are greater than the other coefficients, suggesting that when HQ analysts are
comparatively more informed and, thus, their recommendations deviate more from the consensus,
investors can benefit more from following the mean of HQ analysts’ recommendation revisions.
To estimate the economic magnitude of this market return predictability, we conduct an out-of-
sample trading strategy analysis, reported in the last line of Panel A. The long and short trading
signals are based on the historical variation in monthly mean recommendation changes as follows:
if the average recommendation revision for a given month is greater (smaller) than the median of
the monthly average recommendation revisions over the previous 24 months, i.e., the current
recommendation revision is more optimistic (pessimistic) than they were in the recent past,
investors should buy (short) the market value-weighted index and hold it for one month.20 For the
variable based on the difference between HQ and consensus mean recommendation revisions in
the last column of Panel A, one should buy (short) the market when this variable is positive
(negative), i.e., when HQ analysts, on average, upgrade and are more optimistic (downgrade and
are more pessimistic) than the consensus mean revision. We benchmark the performance of these
20 The results are unaffected by selecting a longer historical window of up to five years. The 24-month window that
we use minimizes the number of months lost to initialize this out-of-sample analysis, while providing enough
observations to robustly calculate the median of monthly mean recommendation changes.
26
strategies against the market return and report the market model alphas. Only HQ analysts’
recommendations produce statistically significant alphas, with a particularly large alpha based on
HQ analysts’ mean recommendation when it is different from the consensus recommendation, at
0.57% per month (6.84% annualized). These findings suggest that investors can earn economically
meaningful returns if they follow the recommendations of a subset of HQ analysts and not the
consensus recommendation.
In Panel B of Table 8, we examine another market-level aggregate relation—whether HQ
analysts’ forecast dispersion aggregated across all firms is associated with systematic uncertainty
in the economy captured by current or future market volatility. We use the VIX to measure
expected market volatility. To capture the degree to which HQ analysts’ uncertainty is different
from the consensus (all analysts’) uncertainty, we define the dispersion ratio for HQ analysts as
the HQ analysts’ forecast dispersion divided by the forecast dispersion of all analysts following
the firm. The dispersion ratios are analogous to the measures of uncertainty in the PEAD analysis.
A bonus feature of HQ and LQ analysts’ forecast dispersions being normalized in the dispersion
ratios is that they are made comparable across firms and years and can thus be suitably aggregated
to measure market uncertainty. A market-level measure of analyst uncertainty is created by firm
value-weighting the dispersion ratios across firms each month, and we also include the results for
aggregate dispersions for completeness.
Panel B reports the regressions of monthly VIX returns (this month’s VIX divided by the
last month’s VIX, minus one) on the dispersions of all, HQ, and LQ analysts’ forecasts and the
HQ and LQ analysts’ dispersion ratios measured before the earnings announcement in the previous
month.21 The results indicate that a higher dispersion ratio for HQ analysts predicts a greater VIX
return in the following month. The LQ dispersion ratio does not capture the level of uncertainty in
21 We also obtain very similar results using the VIX values in lieu of the returns.
27
the market, nor do the non-normalized forecast dispersion variables. To estimate the opportunity
cost for investors not recognizing the limitation of relying on the consensus forecast dispersion,
the last line of Panel B reports the performance of a long-short strategy for the VIX identical in
design to the long-short strategy for the recommendations market return index. If the forecast
dispersion measure in the corresponding column in a given month is above the median of its
monthly values during the previous 24 months, then the strategy is to long the VIX return in the
next month. If the forecast dispersion measure is below the median, then the VIX return is shorted.
To benchmark the performance of these monthly strategy return series, we regress it on the market
return and report the alphas. The only statistically significant (p-value<0.01) abnormal return is
for the trading signal based on HQ analysts’ dispersion ratio; the strategy yields an economically
significant return of 3.32% per month.22 We conclude that when HQ analysts are, on average, more
uncertain than other analysts about the prospects of the firms they follow, investors should expect
an increase in market-wide uncertainty.
Together, consistent with the firm-level findings, the aggregate results indicate that the
market’s fixation on the consensus recommendation and the consensus forecast dispersion results
in a delayed incorporation of information into prices, both at the individual firm level and at the
market level. Investors who recognize quality differences among analysts can benefit from
exploiting this market inefficiency.
7. The definition of high-quality analysts
22 While this alpha is much higher than the typical alphas for stock-based long-short strategies, it is credible because
the market may not be able to fully arbitrage it away. The transaction costs of implementing a VIX strategy with
complete end-of-month rebalancing using VIX futures would be much higher than transaction costs for stocks because
of a significant contango in the VIX futures market, resulting in costs of approximately 14% per month (Nadig, 2016).
Further, the VIX strategy’s average monthly raw return and volatility are 2.3% and 19.6%, respectively.
28
Based on the combined findings of sections 4-6, we reject the null of the economic
efficiency of consensus fixation using a simple measure of analyst quality differences. In this
section, our robustness analysis examines the sensitivity of our finding of the inefficiency of
consensus fixation to the definition of HQ analysts. The firm-year definition used throughout this
paper, which splits analysts into two groups at the median based on the accuracy of their closest
estimate to the annual EPS announcement, is only one of many ways of ranking analysts and
generating an alternative to the consensus forecast. A ranking that, like ours, is based on the last
year’s forecast accuracy is used, for instance, in Loh and Mian (2006) and supported by Sinha,
Brown, and Das (1997), who find it to be superior to rankings based on more years, and Carpenter
and Lynch (1999), who find it to be relatively less exposed to survivorship bias.23 Other ranking
methods can include such variations as using different forecast accuracy cutoffs between the two
groups and giving different weights to forecasts.24 We note that our analysis of the inefficiency of
consensus fixation does not require finding the subset of analysts who beat the consensus by the
23 More generally, rankings based on past performance are common for analyst forecast persistence studies (Stickel,
1992; Sinha, Brown, and Das, 1997) and in a number of other areas, such as the mutual fund and pension fund
performance forecasting (Hendricks, Patel, and Zeckhauser, 1993; Carhart, 1997; Tonks, 2005) and economic
forecasting (Aiolfi and Timmerman, 2006) literatures. Brown (2004) finds that models built on past forecasting
performance predict analysts’ forecasting accuracy as well as a model based on analysts’ individual characteristics,
such as that in Clement (1999). 24 A somewhat more different approach is to rank analysts in a given year by averaging their forecast errors across the
firms they follow. This method underlies various “star” analyst rankings, which are cross-firm by design. This
alternative ranking procedure would avoid losing the observation of the first year when an analyst begins covering a
firm, as we could use the analyst’s ranking in other firms. However, the cross-firm ranking approach has several
pitfalls. First, an aggregated ranking across firms can be misleading if an analyst’s superiority is mainly firm- or
industry-specific (e.g., Kadan et al., 2012). Second, with the cross-firm ranking, we end up with some firms followed
almost exclusively by either HQ or LQ analysts and even populated by just one analyst quality type, coinciding with
the consensus, which would undermine our study’s objective because it relies on comparing the average outputs of
different analyst types to the consensus in each firm. While the cross-firm ranking approach is not suitable for this
study, its ranking characteristics are effectively provided by our firm-year ranking measure: Table 1 shows that an
analyst’s forecast accuracy in a given firm and the accuracy measure based on all firms covered by the analyst are
highly correlated.
29
greatest margin or whose other outputs result in the greatest additional returns relative to the
returns based on the consensus measure.25
The first two alternative ranking procedures that we consider in this section define HQ
analysts as those in the top 70% (with the HQ/LQ proportion at 70%/30%) and the top 30% (with
the HQ/LQ proportion at 30%/70%) of forecast accuracy distribution. Appendix A examines a
broad range of possible cutoff values for HQ analysts and finds that analyst forecasting
performance is persistent over time for all cutoffs. The third alternative definition for HQ and LQ
analysts is computed using all forecasts, rather than the most recent forecast, and gives different
weights to forecasts, depending on how long the forecast has been outstanding. The details of this
time-weighted analyst quality measure are also provided in Appendix A.
Next, we repeat the key firm-level analyses of the paper with the three alternative quality
definitions. If HQ is defined as the top 70% of analysts, then the smallest number of HQ analysts
following the firm is three for their average forecast accuracy to be superior to the consensus
forecast. In that subsample, the market reaction to the consensus forecast is greater than that for
HQ analysts. Importantly, the market does not sufficiently recognize analyst quality differences
and is fixated on the consensus forecast, resulting in predictable mispricing at the earnings
announcement day. These results for forecasts are also robust to defining the top 30% of analysts
as HQ. The results on return and volatility predictability for recommendations and forecast
dispersion are little affected by the first two alternative definitions.26 With the time-weighted
analyst quality definition, HQ analysts’ average forecast is more accurate than the consensus when
four or more analysts follow the firm. The design of the time-weighted measure results in a
25 For these reasons, this study is not related to a literature in economics and statistics aiming to achieve better
forecasting performance than the consensus forecast by proposing different methods of combining forecasts (e.g.,
Conroy and Harris, 1987; Clemen, 1989; Brown, Gay, and Turac, 2008). 26 For HQ defined as the most accurate 70% of analysts, firms with three or more HQ analysts constitute approximately
2/3 of all firms, indicating that our findings about the inefficiency of consensus fixation are not limited to large firms.
30
relatively small sample size, which could explain that while the market reacts more to the
consensus than to the average of HQ analysts’ forecasts based on this measure, this difference is
not statistically significant. However, this result does not mean that the market sufficiently
recognizes analyst quality differences because it would then react more strongly to the average of
HQ forecasts than it would to the consensus. The replication results of the other tables hold for the
time-weighted quality measure as consistently as for the first two alternative cutoff measures.
Overall, we conclude that the definition of HQ analysts and the smallest required number of HQ
analysts associated with this definition are immaterial for our conclusion about the inefficiency of
consensus fixation. More generally, any alternative to the consensus forecast can work for this
analysis as long as it is more accurate than the consensus forecast.
8. Conclusion
Investors can benefit from disregarding the consensus and using a signal with greater
information content. However, the market does not demonstrate an appropriate awareness of these
significant deficiencies in the consensus: the market reacts more strongly to deviations from the
consensus earnings forecast than it does to deviations from HQ analysts’ forecasts. In addition,
HQ analysts’ stock recommendations and forecast dispersion predict the first two moments of firm
and stock market returns better than the corresponding consensus recommendation and forecast
dispersion of all analysts. In short, the persistence of analysts’ differential ability along multitude
dimensions is not sufficiently recognized by the market, resulting in inefficient pricing after
earnings announcements and stock recommendations changes. Overall, our findings suggest that
the market’s fixation on the consensus forecast is not justified. The forecasts and other information
outputs of HQ analysts demonstrably provide superior, more accurate information than the
consensus figures.
31
We show that sell-side research analysts’ superior forecasting ability is reflected not only
in individual stocks but also at the aggregate market level. This result is particularly surprising, as
the limits to arbitrage are much narrower at that level because the aggregate stock market is much
more liquid than individual stocks. This finding implies that the fixation on the consensus is not
limited to individual investors but affects institutional investors, as well. This conclusion is
corroborated by our finding that firms with a greater number of HQ analysts are larger and have
greater analyst following, which are associated with greater institutional holdings. Consequently,
the question arises whether investors’ fixation on the consensus can be remedied. Tracking analyst
performance by the investing public is a necessary step towards this objective. News and
information provider outlets, both electronic and non-electronic, can also play a major role in this
regard by changing how they report and present data on analysts’ expectations to help investors
circumvent their cognitive constraints and fully take into account the variation in analyst quality.
32
Appendix A
The appendix provides details on how several alternative classifications of analysts into
the HQ and LQ groups affect the persistence in analyst forecasting performance. The ranking
procedure sorts analysts in a given firm-year based on their absolute forecast error. In general, HQ
analysts are those who are ranked in the top 𝑝 percent of analysts, while LQ analysts are those in
the bottom (1 − 𝑝) percent. If analysts’ forecasting performance were uncorrelated across years,
then the fractions of analysts who preserve their ranking in two consecutive years as HQ and LQ
would be 𝑝2 and (1 − 𝑝)2, respectively, or 𝑝2 + (1 − 𝑝)2 of all analysts.
Figure A.1 plots the fraction of analysts who retain their rankings in consecutive years and
the expected fraction assuming no performance correlation across years. In this figure, to rank
analysts up to the decile on the horizontal axis, the sample is constrained to firms that are followed
by ten or more analysts. We find that with all cutoff values of 𝑝 on the horizontal axis, the actual
fraction of persistent forecasting performance is above the expected fraction and that all the
differences are statistically significant (p-value<0.01). For example, when we classify the top 10%
of analysts following a firm in a given year as HQ (p=10%) and the bottom 90% as LQ, the
expected fraction given random assignment is 0.92 + 0.12 = 0.82. The figure shows that the
actual fraction is greater than that at 0.843. The overall finding is that for all of the cutoff values,
there is a sizeable persistent component; thus, for accuracy persistence, which exact cutoff value
we choose to partition HQ and LQ analysts should make little difference.
The two alternative cutoffs for HQ analysts—the top 70% (so that the LQ analysts are the
bottom 30%) of analysts and the top 30% (the LQ analysts are the bottom 70%) of analysts, which
are considered in Section 7—are symmetric around the cutoff at the median used throughout this
paper. Because analysts are ranked in year t-1, the proportion of HQ and LQ analysts following
the firm tends to become different at the year t announcement, potentially resulting in too few HQ
or LQ analysts for the firm. To avoid small sample bias, these alternative analyst quality definitions
33
require the same restrictions on the sample as those used in the rest of this paper, with an additional
condition. For the robustness analysis, we require that the proportion of HQ and LQ analysts does
not change by more than a 20% margin from t-1 to t. The motivation is obvious: for the definition
of HQ analysts as the top 30% in year t-1, the additional condition ensures that their fraction always
stays above 10% of analysts covering the firm in year t; and if HQ analysts are defined as the top
70% in year t-1, then their fraction can be between 50% and 90% of analysts in year t, while the
fraction of LQ analysts never falls below 10% of analysts covering the firm in year t.
The third ranking is based on the time-weighted absolute forecast error, which is computed
using all forecasts by the analyst during the 300 days prior to the annual earnings announcement.
This measure is similar to the time-weighted bid-ask spread that is commonly used in the market
microstructure literature (e.g., McInish and Wood, 1992; Lee, Mucklow, and Ready, 1993;
Bessembinder, 1999; Rubin, 2007). For a given announcement and analyst, it is computed as
follows:
𝑇𝑊𝐹𝐸𝑡 =𝐹𝐸300×𝑑1+∑ (𝐹𝐸𝑗×𝑑𝑗)𝑛
𝑗=2
300 (3)
where 𝑇𝑊𝐹𝐸𝑡 is the time-weighted absolute forecast error of the analyst in year t; 𝐹𝐸300 is the
absolute forecast error based on the forecast outstanding on the 300th day prior to the earnings
announcement; 𝑑1 is the number of days that this forecast is outstanding (from the 300th day prior
to the earnings announcement to the earliest of the earnings announcement day or the following
earnings forecast revision day); n-1 is the number of estimates issued by the analyst between the
299th day prior to the earnings announcement and the earnings announcement day; 𝐹𝐸𝑗 is the
absolute forecast error of forecast j; and 𝑑𝑗 is the number of days that the forecast has been
outstanding. The advantage of the time-weighted measure is that it captures the analyst’s ability
over a four-quarter period, instead of at a single point just before the annual earnings
announcement. However, the measure’s disadvantage is that it excludes analysts who have not
34
provided an annual forecast 300 or more days prior to the annual earnings announcement day. The
sample therefore shrinks by approximately 65% relative to the sample used throughout this paper,
thus both reducing the power of our empirical analysis with this quality measure and possibly
biasing the results because it remains unknown whether the lack of early forecasts is due to the
analyst’s poor ability or a neutral reason, such as common practice in the given industry or firm
being analyzed. These disadvantages are among the reasons the analyst classification used in our
paper equal-weights the forecasts made by the best performing subset of analysts, which is
certainly more common in the literature and analogous to the approach examined, for instance, in
Aiolfi and Timmerman (2006). Further, studies focused on the forecast weighting topic find that
the simple averaging of expert forecasts is commonly equivalent or more optimal than more
sophisticated weighting methods for various economic series (Genre et al., 2013).
35
Figure A.1: Persistence in analysts’ forecasting performance. This figure depicts how the fraction of
analysts retaining their ranking of either high- or low-quality (HQ and LQ) in terms of forecast accuracy in
two consecutive years depends on the cutoff percentile in the definition of HQ analysts. HQ analysts are
those whose closest absolute forecast errors are less than the absolute forecast error at the cutoff percentile
(the horizontal axis) of the distribution of forecast errors for the firm’s annual earnings announcement in
year t-1. The closest absolute forecast error is the absolute difference between an analyst’s forecast estimate
closest to the earnings announcement prior to the announcement day and the actual annual earnings, divided
by the share price at the beginning of the calendar year. Expected performance assuming no persistence is
the fraction of analysts who would have the same forecast performance category in two consecutive years
if their performance were uncorrelated between years.
45%
50%
55%
60%
65%
70%
75%
80%
85%
90%
10% 15% 20% 25% 30% 35% 40% 45% 50% 55% 60% 65% 70% 75% 80% 85% 90%
Perc
enta
ge o
f analy
sts
with t
he s
am
e q
ualit
y cate
gory
in t
wo c
onsecutive
years
Percentage of analysts defined as HQ
Expected performance (assuming no persistence) Actual performance
36
References
Abarbanell, J.S., Lanen, W.N. Verrecchia, R.E., 1995. Analysts' forecasts as proxies for investor
beliefs in empirical research. Journal of Accounting and Economics, 20(1), 31-60.
Aiolfi, M., Timmermann, A., 2006. Persistence in forecasting performance and conditional
combination strategies. Journal of Econometrics, 135(1), 31-53.
Ajinkya, B.B., Gift, M.J., 1985. Dispersion of financial analysts' earnings forecasts and the (option
model) implied standard deviations of stock returns. The Journal of Finance, 40(5), 1353-
1365.
Altinkiliç, O. and Hansen, R.S., 2009. On the information role of stock recommendation revisions.
Journal of Accounting and Economics, 48(1), 17-36.
Babenko, I., Tserlukevich, Y., Vedrashko, A., 2012. The credibility of open market share
repurchase signaling. Journal of Financial and Quantitative Analysis, 47(5), 1059-1088.
Baker, S.R., Bloom, N. and Davis, S.J., 2016. Measuring economic policy uncertainty. The
Quarterly Journal of Economics, 131(4), 1593-1636.
Barber, B., Lehavy, R., McNichols, M., Trueman, B., 2001. Can investors profit from the
prophets? Security analyst recommendations and stock returns. The Journal of Finance,
56(2), 531-563.
Barron, O.E., Stanford, M.H. and Yu, Y., 2009. Further evidence on the relation between analysts'
forecast dispersion and stock returns. Contemporary Accounting Research, 26(2), 329-357.
Bessembinder, H., 1999. Trade execution costs on Nasdaq and the NYSE: A post-reform
comparison. Journal of Financial and Quantitative Analysis, 34(3), 387-407.
Bradshaw, M.T., 2004. How do analysts use their earnings forecasts in generating stock
recommendations? The Accounting Review, 79(1), 25-50.
Brown, L.D., 2004. How important is past analyst forecast accuracy? Financial Analysts Journal
57, 44-49.
Brown, L.D., Gay, G.D. and Turac, M., 2008. Creating a “smart” conditional consensus forecast.
Financial Analysts Journal, 64(6), 74-86.
Buraschi, A., Piatti, I., Whelan P., 2017. Expected Term Structures, University of Chicago working
paper.
Carhart, M.M., 1997. On persistence in mutual fund performance. The Journal of Finance, 52(1),
57-82.
Carpenter, J.N., Lynch, A.W., 1999. Survivorship bias and attrition effects in measures of
performance persistence. Journal of Financial Economics, 54(3), 337-374.
Chiang, Chin-Han; Fan, Jianqing; Hong, Harrison; Tu, Jun, 2018. Robust measures of earnings
surprises. Journal of Finance, forthcoming.
Clemen, R.T., 1989. Combining forecasts: A review and annotated bibliography. International
Journal of Forecasting, 5(4), 559-583.
37
Clement, M.B., 1999. Analyst forecast accuracy: Do ability, resources, and portfolio complexity
matter? Journal of Accounting and Economics, 27(3), 285-303.
Consensus earnings estimates, Special 12-page report, April 2013, IR Magazine,
www.InsideInvestorRelations.com
Conroy, R. and Harris, R., 1987. Consensus forecasts of corporate earnings: Analysts' forecasts
and time series methods. Management Science, 33(6), 725-738.
Cooper, R.A., Day, T.E. and Lewis, C.M., 2001. Following the leader: a study of individual
analysts’ earnings forecasts. Journal of Financial Economics, 61(3), 383-416.
Da, Z., Engelberg, J. and Gao, P., 2015. The sum of all fears investor sentiment and asset prices.
Review of Financial Studies, 28(1), 1-32.
De Franco, G., Zhou, Y., 2009. The performance of analysts with a CFA designation: the role of
human-capital and signaling theories. The Accounting Review, 84(2), 383-404.
Diether, K.B., Malloy, C.J., Scherbina, A., 2002. Differences of opinion and the cross section of
stock returns. The Journal of Finance, 57(5), 2113-2141.
Ertimur, Y., Sunder, J., Sunder, S.V., 2007. Measure for measure: The relation between forecast
accuracy and recommendation profitability of analysts. Journal of Accounting Research,
45(3), 567-606.
Fama, E.F., French, K.R., 1993. Common risk factors in the returns on stocks and bonds. Journal
of Financial Economics, 33(1), 3-56.
Fang, L.H. and Yasuda, A., 2014. Are stars’ opinions worth more? The relation between analyst
reputation and recommendation values. Journal of Financial Services Research, 46(3), 235-
269.
Francis, J., Lafond, R., Olsson, P., Schipper, K., 2007. Information uncertainty and post‐earnings‐announcement‐drift. Journal of Business Finance and Accounting, 34(3‐4), 403-433.
Francis, J. and Soffer, L., 1997. The relative informativeness of analysts' stock recommendations
and earnings forecast revisions. Journal of Accounting Research, 35(2), 193-211.
Glushkov, D., Robinson, D., 2006. A Note on IBES Unadjusted Data, WRDS Documentation on
IBES, https://wrds-
web.wharton.upenn.edu/wrds/support/Data/_001Manuals%20and%20Overviews/_003I-B-
E-S/index.cfm.
Genre, V., Kenny, G., Meyler, A., Timmermann, A., 2013. Combining expert forecasts: can
anything beat the simple average? International Journal of Forecasting, 29(1), 108-121.
Hendricks, D., Patel, J., Zeckhauser, R., 1993. Hot hands in mutual funds: Short‐run persistence
of relative performance, 1974–1988. The Journal of Finance, 48(1), 93-130.
Hirshleifer, D., Lim, S., Teoh, S. H., 2009. Driven to distraction: Extraneous events and
underreaction to earnings news. Journal of Finance 64, 2289-2325.
38
Hirst, D.E., Hopkins, P.E., Wahlen, J.M., 2004. Fair values, income measurement, and bank
analysts' risk and valuation judgments. The Accounting Review, 79(2), 453-472.
Hung, M., Li, X., Wang, S., 2014. Post-earnings-announcement drift in global markets: Evidence
from an information shock. Review of Financial Studies, 28 (4), 1242-1283.
Hutton, A.P., Lee, L.F., Shu, S.Z., 2012. Do managers always know better? The relative accuracy
of management and analyst forecasts. Journal of Accounting Research, 50(5), 1217-1244.
Jacob, J., Lys, T.Z., Neale, M.A., 1999. Expertise in forecasting performance of security
analysts. Journal of Accounting and Economics, 28(1), 51-82.
Jegadeesh, N., Kim, J., Krische, S.D., Lee, C., 2004. Analyzing the analysts: When do
recommendations add value? The Journal of Finance, 59(3), 1083-1124.
Johnson, T. 2004. Forecast dispersion and the cross-section of expected returns. The Journal of
Finance 59 (5): 1957–78.
Kadan, O., Madureira, L., Wang, R. and Zach, T., 2012. Analysts' industry expertise. Journal of
Accounting and Economics, 54(2-3), 95-120.
Kirk, M. P., Reppenhagen, D.A., Tucker, J.W., 2014. Meeting individual analyst expectations,
Accounting Review 89, 2203–2231.
Krishnaswami, S., Subramaniam, V., 1999. Information asymmetry, valuation, and the corporate
spin-off decision. Journal of Financial Economics, 53(1), 73-112.
Kucheev, Y.O., Ruiz, F. and Sorensson, T., 2017. Do Stars Shine? Comparing the Performance
Persistence of Star Sell-Side Analysts Listed by Institutional Investor, the Wall Street
Journal, and StarMine. Journal of Financial Services Research, 52(3), 277-305.
Lee, C.M., Mucklow, B. and Ready, M.J., 1993. Spreads, depths, and the impact of earnings
information: An intraday analysis. Review of Financial Studies, 6(2), 345-374.
Li, Y., Ng, D.T., Swaminathan, B., 2013. Predicting market returns using aggregate implied cost
of capital. Journal of Financial Economics, 110(2), 419-436.
Lim, T., 2001. Rationality and analysts' forecast bias. The Journal of Finance, 56(1), 369-385.
Lin, H.W., McNichols, M.F., 1998. Underwriting relationships, analysts' earnings forecasts and
investment recommendations. Journal of Accounting and Economics, 25(1), 101-127.
Loh, R.K., Mian, G.M., 2006. Do accurate earnings forecasts facilitate superior investment
recommendations? Journal of Financial Economics, 80(2), 455-483.
Loh, R.K., Stulz, R.M., 2011. When are analyst recommendation changes influential? Review of
Financial Studies, 24(2), 593-627.
Ludvigson, S.C., 2004. Consumer confidence and consumer spending. Journal of Economic
Perspectives, 18(2), 29-50.
Maines, L.A., McDaniel, L.S., Harris, M.S., 1997. Implications of proposed segment reporting
standards for financial analysts' investment judgements. Journal of Accounting
Research, 35, 1-24.
39
Malloy, C.J., 2005. The geography of equity analysis. Journal of Finance, 60(2), 719-755.
McInish, T.H. and Wood, R.A., 1992. An analysis of intraday patterns in bid/ask spreads for NYSE
stocks. Journal of Finance, 47(2), pp.753-764.
Mendenhall, R.R., 2004. Arbitrage risk and post-earnings-announcement drift. Journal of Business
77, 875-894.
Michaely, R., Womack, K.L., 1999. Conflict of interest and the credibility of underwriter analyst
recommendations. Review of Financial Studies, 12(4), 653-686.
Mikhail, M.B., Walther, B.R., Willis, R.H., 1997. Do security analysts improve their performance
with experience? Journal of Accounting Research, 35, 131-157.
Nadig, D., 2016. Understanding VIX ETFs: Careful What You Wish For.
http://www.etf.com/sections/blog/understanding-vix-etfs, September 2.
Novy-Marx, R., Velikov, M. 2016. A taxonomy of anomalies and their trading costs. Review of
Financial Studies 29, 104-147.
Rubin, A., 2007. Ownership level, ownership concentration and liquidity. Journal of Financial
Markets, 10(3), 219-248.
Rubin, A., Segal, B., Segal D., 2017. The interpretation of unanticipated news arrival and analysts'
skill, Journal of Financial and Quantitative Analysis 52(4), 1491-1518.
Sadka, R., 2006. Momentum and post-earnings-announcement drift anomalies: The role of
liquidity risk. Journal of Financial Economics 80, 309-349.
Sinha, P., L. Brown, Das S., 1997. A re-examination of financial analysts' differential earnings
forecast ability. Contemporary Accounting Research 14, 1-42.
So, E., 2013. A new approach to predicting analyst forecast errors: Do investors overweight analyst
forecast? Journal of Financial Economics 108, 615-640.
Sorescu, S., Subrahmanyam, A., 2006. The cross section of analyst recommendations. Journal of
Financial and Quantitative Analysis, 41(1), 139-168.
Stickel, S., 1992. Reputation and performance among security analysts. Journal of Finance 47,
1811-36.
Tatler, B.W., 2007. The central fixation bias in scene viewing: Selecting an optimal viewing
position independently of motor biases and image feature distributions. Journal of Vision,
7(14):4, 1-17.
Tonks, I., 2005. Performance persistence of pension‐fund managers. The Journal of Business,
78(5), 1917-1942.
Wahlen, J.M. and Wieland, M.M., 2011. Can financial statement analysis beat consensus analysts’
recommendations? Review of Accounting Studies, 16(1), 89-115.
Yezegel, A., 2015. Why do analysts revise their stock recommendations after earnings
announcements? Journal of Accounting and Economics, 59(2), 163-181.
40
Zhang, X., 2006a. Information uncertainty and stock returns. Journal of Finance, 61(1), 105-137.
Zhang, X., 2006b. Information uncertainty and analyst forecast behavior. Contemporary
Accounting Research, 23(2), 565-590.
41
Figure 1: Absolute forecast errors up to 300 days before the annual earnings announcement day. The absolute forecast error is the absolute
difference between an analyst’s forecast estimate and the actual annual earnings, divided by the share price at the beginning of the calendar year. The absolute forecast error at day -t prior to the earnings announcement date is calculated as the mean absolute forecast error based on all annual
earnings forecasts outstanding for the firm as of that day, averaged across the firm’s announcements and then averaged across firms. HQ analysts
are those whose closest absolute forecast errors are below the median closest absolute forecast error for the earnings announcement in the previous
year, where the closest absolute forecast error is that closest to the announcement day prior to that day.
0.010
0.011
0.012
0.013
0.014
0.015
0.016
0.017
0.018
0.019
0.020
0.021
0.022
0.023
0.024
0.025
0.026
0.027
-300 -280 -260 -240 -220 -200 -180 -160 -140 -120 -100 -80 -60 -40 -20 0
Mean
ab
so
lute
fo
recast
err
or
Days before earnings announcement
All analysts High quality analysts
42
Figure 2: Cumulative post-announcement drifts and analysts’ uncertainty This figure shows the cumulative returns for 1- to 11-month horizons following earnings announcements. Each month, stocks enter a calendar-time
long (short) portfolio, depending on whether their earnings surprise is positive (negative), where earnings surprise is defined based on the consensus
estimate in Table 2. The horizontal axis is the drift’s horizon, which is the number of months a stock is held in the calendar-time portfolio. The
monthly long-minus-short value-weighted portfolio return is regressed on the four Fama-French-Carhart factors. The cumulative drift for a given
drift horizon on the vertical axis is calculated as the regression intercept (monthly alpha) multiplied by the portfolio’s horizon length in months. The
graphs are shown for the full sample and two subsamples of firms in which the standard deviation of HQ analysts’ forecasts (SD HQ) is either greater
or smaller than that of all analysts’ forecasts (SD consensus) for the firm-year. Analysts are ranked as HQ based on their previous year’s absolute
forecast errors, defined in Table 1.
-4
-2
0
2
4
6
8
10
12
14
0 1 2 3 4 5 6 7 8 9 10 11
Cum
ula
tive r
etu
rn (
%)
Horizon (months)
Full sample SD HQ > SD consensus SD HQ < SD consensus
43
Table 1: Individual analyst analysis: characteristics and persistence of forecast accuracy Panel A conducts univariate analysis for high- and low-quality (HQ and LQ) analysts. HQ (LQ) quality analysts are those
whose closest absolute forecast errors are below (at or above) the median closest absolute forecast error for the firm’s
earnings announcement. The closest absolute forecast error for an analyst is the absolute difference between the analyst’s
forecast closest to the earnings announcement prior to the announcement day and the actual annual earnings, divided by
the share price at the beginning of the calendar year. Overall tenure is the number of years since the analyst first appeared
in the I/B/E/S file. Firm-specific tenure is the number of years since the analyst began covering the specific firm in the
I/B/E/S file. Brokerage house size is the number of analysts in the analyst’s brokerage house. Firms covered is the number
of firms covered by the analyst. Panel B shows probit models predicting the HQ status indicator, which equals one if the
analyst is ranked as HQ in a given firm in the current year and zero otherwise (columns (1), (2), (5), and (6)) and the
regressions for the analyst’s closest absolute forecast error in a given firm in the current year in columns (3) and (4). HQ
status in other firms equals one (zero) if the analyst is of HQ in the majority of the other firms that the analyst follows
during the year; analysts who have an equal number of HQ and LQ rankings in other firms are excluded. Firm size is the
log of the firm’s market value of equity, equal to the stock price times the number of shares outstanding at the end of the
month prior to the annual earnings announcement. Annual return is the annual return of the firm's equity over the 12
months prior to the earnings announcement month. Leverage is the book value of total liabilities divided by the book
value of total assets, and Book-to-market is the book value of common equity divided by the market value of equity at
the end of the fiscal year. Number of analysts is the number of analysts following the firm. All independent variables are
measured prior to the announcement date. The probit coefficients are reported as marginal probability effects. All models
include the intercept. Robust standard errors are clustered by firm. z-statistics for probit models and t-statistics elsewhere
are in parentheses. *, **, and *** indicate significance at the 10%, 5%, and 1% level, respectively.
Panel A. HQ and LQ analyst characteristics
Analyst or announcement
characteristic HQ analysts LQ analysts
Difference
(t-statistic)
Overall tenure 7.07 7.00 0.07*** (4.73)
Firm-specific tenure 3.04 2.97 0.07*** (8.61)
Brokerage house size 65.76 63.04 2.72*** (19.14)
Firms covered 17.60 17.55 0.05* (1.79)
Absolute forecast error 0.0081 0.0089 -0.008*** (-12.83)
Forecast error 0.00185 0.00181 0.00004 (0.66)
44
Panel B. Persistence of analysts’ forecasting performance
HQ status Absolute forecast error HQ status
(1) (2) (3) (4) (5) (6)
HQ status (year t-1) 0.0414*** 0.0407*** -0.0007*** -0.0007*** 0.0531*** 0.0382***
(25.54) (25.13) (-16.14) (-15.91) (32.95) (22.84)
HQ status in other firms (year t-1) 0.0387*** 0.0514***
(23.14) (31.83)
Firm size 0.0034*** 0.0011*** -0.0061*** -0.0061***
(10.52) (3.11) (-25.12) (-25.15)
Annual return -0.0003 0.0005 -0.0009*** -0.0009***
(-0.50) (0.88) (-5.68) (-5.67)
Leverage 0.0006 0.0003 0.00573*** 0.0057***
(0.37) (0.20) (6.58) (6.55)
Book-to-market 0.0001 0.00003 0.00002 0.00002
(1.40) (0.54) (1.48) (1.48)
Number of analysts 0.0007*** 0.0007*** 0.00022*** 0.0002***
(12.40) (14.52) (10.78) (10.69)
Overall tenure 0.0007*** -0.00004*** 0.0002 (4.33) (-7.12) (1.12)
Firm-specific tenure 0.0023*** 0.00001 0.003***
(9.66) (1.45) (12.90)
Brokerage house size 0.0001*** -0.000001 0.0001***
(10.63) (-0.26) (7.23)
Firm coverage -0.0007*** 0.00003*** -0.0005***
(-12.32) (8.40) (-8.63)
Year fixed effects Yes Yes Yes Yes Yes Yes
Firm fixed effects Yes Yes
Observations (analyst-years) 485,815 485,815 485,815 485,815 443,262 443,262
Adj. R-squared 0.344 0.344
45
Table 2: The number of HQ analysts and the accuracy of their forecasts relative to the consensus Panel A compares the accuracy of the HQ analysts’ average forecast and the consensus forecast sorted by the number of
HQ analysts following the firm in a given earnings announcement year. Analysts are ranked as HQ based on their previous
year’s absolute forecast errors, defined in Table 1. The SUE of consensus (SUE of HQ analysts) is standardized
unexpected earnings, equal to the difference between the actual earnings and the average forecast provided by all analysts
(HQ analysts) following the firm normalized by the stock price at the beginning of the year. Absolute SUE is the absolute
value of SUE. Accuracy improvement is the percentage reduction from the absolute SUE of consensus to the absolute
SUE of HQ analysts. t-statistics in the last column are for the difference in the means test between the absolute SUEs of
the consensus and HQ analysts. Panel B reports sample characteristics, defined in Table 1, for two samples sorted based
on whether the firm is followed by fewer than four HQ analysts or four or more HQ analysts in a given announcement
year. *, **, and *** indicate significance at the 10%, 5%, and 1% level, respectively.
Panel A. The smallest number of HQ analysts needed to improve over the consensus forecast
Number of
HQ analysts
Absolute SUE of
Consensus
Absolute SUE of
HQ analysts
Accuracy
improvement
t-statistics
Abs. SUE
difference
1 or more 0.00656 0.00678 -3.31% -8.63***
2 or more 0.00589 0.00595 -1.08% -3.19***
3 or more 0.00514 0.00513 0.19% 0.54
4 or more 0.00461 0.00455 1.17% 2.99***
5 or more 0.00422 0.00415 1.52% 3.51***
6 or more 0.00404 0.00396 1.96% 3.95***
7 or more 0.00386 0.00377 2.35% 4.47***
8 or more 0.00377 0.00367 2.69% 4.60***
9 or more 0.00355 0.00346 2.61% 3.91***
10 or more 0.00346 0.00337 2.59% 3.44***
46
Panel B: Comparing samples in which HQ analysts’ average forecast is either less or more
accurate than the consensus
Difference
(t-statistics)
Number of HQ analysts
4 or more Fewer than 4
Firm characteristics:
9,149*** (43.78) 10,522 1,372 Firm size ($M)
10.4*** (210.1) 16.7 6.3 Number of analysts
-0.67*** (-6.49) 0.66 1.34 Book-to-market
0.04** (14.93) 0.57 0.53 Leverage
-1.3* (-1.74) 20.4 21.7 Annual return (%)
891 1136 Avg. number of firms per year
Analyst characteristics:
0.51*** (8.66) 7.42 6.91 Overall tenure
1.10*** (29.48) 3.43 2.33 Firm-specific tenure
12.3*** (18.73) 69.4 57.1 Brokerage house size
0.5*** (3.73) 18.2 17.7 Firms covered
47
Table 3: The immediate reaction to earnings announcements This table reports the earnings response coefficients for measures of earnings surprise (SUE) based on the forecasts of all
analysts following the firm (consensus) and on the forecasts of high- and low-quality (HQ and LQ) analysts. Analysts are
ranked as HQ or LQ based on their previous year’s absolute forecast errors, defined in Table 1. The dependent variable is
the buy-and-hold abnormal return (based on the four-factor Fama-French and Carhart model) for the earnings
announcement day and the following trading day. The SUE of Consensus and SUE of HQ and LQ analysts are defined in
Table 2. All other variables are defined in Table 1. Columns (1)-(3) use the entire sample of earnings announcements, and
columns (4)-(6) use the sample of earnings announcements by firms followed by at least four HQ analysts. All independent
variables other than SUE are measured prior to the announcement date. The intercept and year fixed effects are included
in all regressions. Robust standard errors are clustered by firm. t-statistics are provided in parentheses. The last two lines
report the p-values for the chi-squared tests of the equality of the coefficients on SUE measures for the three analyst groups.
*, **, and *** indicate significance at the 10%, 5%, and 1% level, respectively.
Full Sample 4 or more HQ analysts
(1) (2) (3) (4) (5) (6)
SUE of Consensus 0.7245*** 0.7526*** (13.62) (5.20)
SUE of HQ analysts 0.6211*** 0.6927***
(12.93) (5.12)
SUE of LQ analysts 0.5691*** 0.6044*** (12.78) (4.85)
Firm size -0.0003 -0.0002 -0.0001 -0.0007 -0.0007 -0.0007 (-0.71) (-0.55) (-0.37) (-1.19) (-1.15) (-1.14)
Annual return -0.0006 -0.0006 -0.0006 0.0007 0.0007 0.0007 (-0.98) (-0.89) (-0.87) (0.68) (0.72) (0.71)
Leverage 0.0059*** 0.0057*** 0.0057*** 0.0030 0.0029 0.0028 (3.73) (3.55) (3.56) (1.24) (1.19) (1.12)
Book-to-market 0.00002 0.00001 0.00001 -0.0002*** -0.00018*** -0.00023*** (0.38) (0.29) (0.28) (-9.20) (-7.67) (-11.14)
Number of analysts 0.00002 0.00002 0.00001 -0.00004 -0.00003 -0.00004 (0.25) (0.28) (0.08) (-0.42) (-0.40) (-0.45)
Observations 44,709 44,709 44,709 20,221 20,221 20,221
Adj. R-squared 0.015 0.013 0.013 0.011 0.010 0.009
p-value (SUE of HQ
analysts vs. SUE of
consensus)
0.000 0.009
p-value (SUE of HQ
analysts vs. SUE of LQ
analysts)
0.02 0.01
48
Table 4: Predicting market reaction on the earnings announcement day The dependent variable is the buy-and-hold abnormal return (based on the four-factor Fama-French and Carhart model)
for the earnings announcement day and the following trading day. Analysts are ranked as HQ or LQ based on their
previous year’s absolute forecast errors, defined in Table 1. Predicted surprise is equal to (HQ analysts’ average forecast
minus the consensus forecast) in columns (1) and (2) and (HQ analysts’ average forecast minus LQ analysts’ average
forecast) in columns (3) and (4), normalized by the stock price at the beginning of the year. The positive predicted surprise
indicator equals one if predicted surprise is positive and zero if it is negative. Big positive predicted surprise equals one
if predicted surprise is greater than the median of positive values of predicted surprise and zero if predicted surprise is
smaller than the median of the negative values of predicted surprise in year t-1. All independent variables are measured
prior to the announcement date, and the regressions include the intercept and year fixed effects. Robust standard errors
are clustered by firm. t-statistics are provided in parentheses. The last line of the table provides the two-day holding
returns of a trading strategy that is long if the predicted surprise indicator variable in that column is equal to 1 and short
if it is equal to 0. *, **, and *** indicate significance at the 10%, 5%, and 1% level, respectively.
Predicted surprise:
HQ average – Consensus
Predicted surprise:
HQ average – LQ average
(1) (2) (3) (4)
Positive predicted surprise 0.0015* 0.0019**
(1.92) (2.48)
Big positive predicted surprise 0.0007* 0.0008**
(1.71) (1.98)
Firm size 0.0003 -0.00002 0.0003 -0.00002 (0.76) (-0.03) (0.75) (-0.04)
Annual return -0.0002 -0.0006 -0.0002 -0.0010 (-0.25) (-0.67) (-0.24) (-1.15)
Leverage 0.0040** 0.0071*** 0.0040** 0.0072*** (2.45) (2.89) (2.45) (2.90)
Book-to-market 0.00001 0.00001 0.00001 0.00001 (0.03) (0.15) (0.03) (0.11)
Number of analysts -0.00001 -0.00006 -0.00001 -0.00002
(-0.02) (-0.51) (-0.04) (-0.15)
Observations 44,709 20,999 44,709 20,605
Adj. R-squared (%) 0.086 0.078 0.171 0.230
Two-day long-short strategy
returns (%)
0.14*
(1.88)
0.20*
(1.64)
0.19**
(2.52)
0.24*
(1.94)
49
Table 5: Returns following recommendation revisions The dependent variable is the firm’s stock return in the calendar month following the month with a recommendation
revision. The sample consists of all recommendation revisions in the month when the annual earnings is announced by
the firm. A recommendation is an integer from 1 to 5, where 1 is strong buy, 5 is strong sell, and 3 is hold. A
recommendation revision is the negative of the difference between the current and the previous recommendations of an
analyst; thus, a positive (negative) recommendation revision is an upgrade (downgrade). The consensus, HQ, and LQ
recommendation revision variables are the averages of individual analysts’ revisions of all, HQ, and LQ analysts
following the firm, respectively, during the earnings announcement month. Analysts are ranked as HQ or LQ based on
this year’s absolute forecast errors, defined in Table 1. HQ-consensus is equal to HQ analysts’ average recommendation
revision minus the consensus recommendation revision if HQ analysts’ average recommendation revision is either
positive and greater or negative and smaller (more negative) than the consensus recommendation revision; otherwise, this
variable is zero. The other independent variables are defined in Table 1 and measured prior to the earnings announcement.
All regressions include the intercept, and robust standard errors are clustered by firm. The last two lines provide the event-
time raw long-short strategy returns and calendar-time alphas of the value-weighted long-short portfolio based on the
four-factor Fama-French and Carhart model in the month following the month with the revision. The long (short) position
is in the firms for which the mean recommendation revision is positive (negative), according to the consensus, HQ, and
LQ analysts’ recommendation revision variables in columns (1)-(3), respectively. In column (4), the position in the stock
is long (short) if HQ-consensus is positive (negative); otherwise, the stock does not enter either portfolio. t-statistics are
provided in parentheses. *, **, and *** indicate significance at the 10%, 5%, and 1% level, respectively.
(1) (2) (3) (4)
Consensus recommendation revision 0.0015*
(1.67)
HQ analysts’ recommendation revision 0.0025**
(2.09)
LQ analysts’ recommendation revision 0.0004
(0.30)
HQ-consensus 0.0158***
(2.62)
Lagged dependent variable 0.0141 0.0143 0.0158 0.0161
(1.01) (1.04) (1.15) (1.19)
Firm size -0.0043*** -0.0043*** -0.0043*** -0.0043***
(-4.26) (-4.25) (-4.20) (-4.20)
Leverage 0.0008 0.0008 0.0007 0.0008
(0.18) (0.19) (0.16) (0.19)
Book-to-market -0.00001 -0.00001 -0.00001 -0.00001
(-0.34) (-0.30) (-0.32) (-0.32)
Number of analysts 0.0003* 0.0003* 0.0003* 0.0003*
(1.80) (1.81) (1.78) (1.78)
Year fixed effects Yes Yes Yes Yes
Observations 21,381 21,381 21,381 21,381
Adj. R-squared 0.0430 0.0420 0.0419 0.0421
Long-short raw return next month (%) 0.36 0.85*** 0.14 1.24***
(1.61) (3.11) (0.49) (3.34)
Long-short monthly calendar-time alpha (%) -0.16
(-0.45)
0.32
(0.79)
-0.31
(-0.71)
1.06*
(1.80)
50
Table 6: Forecast dispersion and future return volatility The dependent variable is the standard deviation of the firm’s daily returns in the month following the
annual earnings announcement month. Consensus, HQ, and LQ analysts’ forecast dispersion are the
standard deviation of the consensus, HQ, and LQ analysts’ forecasts, respectively, normalized by the
stock price and using each analyst’s closest forecast issued during the 60 days prior to the earnings
announcement. Analysts are ranked as HQ or LQ based on their previous year’s absolute forecast errors,
defined in Table 1. The other independent variables are defined in Table 1 and measured prior to the
announcement date. All regressions include the intercept. Robust standard errors are clustered by firm.
t-statistics are provided in parentheses. *, **, and *** indicate significance at the 10%, 5%, and 1% level,
respectively.
(1) (2) (3)
Consensus dispersion 0.0730
(1.63)
HQ analysts’ dispersion 0.1315**
(2.55)
LQ analysts’ dispersion 0.0340
(0.79)
Lagged dependent variable 0.5179*** 0.5141*** 0.5206***
(12.69) (12.82) (12.63)
Firm size -0.0007 -0.0006 -0.0008*
(-1.59) (-1.40) (-1.73)
Annual return 0.0002 0.0002 0.0002
(0.53) (0.56) (0.46)
Leverage -0.0007 -0.0007 -0.0005
(-0.29) (-0.31) (-0.23)
Book-to-market 0.0015 0.0016 0.0015
(1.25) (1.41) (1.16)
Number of analysts -0.00004 -0.00004 -0.00003
(-1.22) (-1.40) (-1.13)
Year fixed effects Yes Yes Yes
Firm fixed effects Yes Yes Yes
Observations 4,812 4,812 4,812
Adj. R-squared 0.691 0.693 0.691
51
Table 7: Post-earnings announcement drift and analysts’ uncertainty The table reports the monthly abnormal returns for 1- to 11-month horizons following annual earnings
announcements. Announcements are divided into two subsamples in which the standard deviation of HQ
analysts’ forecast errors is greater (the high-uncertainty sample) or smaller (the low-uncertainty sample) than the
standard deviation of forecast errors for all analysts following the firm. Analysts are ranked as HQ based on their
previous year’s absolute forecast errors, defined in Table 1. A stock is assigned to the long or short portfolio,
depending on whether its earnings surprise is positive or negative, respectively, where earnings surprise is defined
based on the consensus estimate in Table 2. A stock enters a portfolio at the beginning of the month following
the month of the announcement and is held for the length of the horizon. In columns (1)-(6), the monthly value-
weighted portfolio returns are regressed on the four Fama-French-Carhart factors to obtain the reported drift,
which is the intercept of the regression (monthly alpha). For the monthly alphas reported in column (7), the high-
uncertainty long-short post-announcement raw return is regressed on the full-sample long-short post-
announcement raw return and the four Fama-French-Carhart factors. *, **, and *** represent 10%, 5%, 1%,
significance based on the regression t-statistics, respectively.
PEAD horizon
(months)
High uncertainty Low uncertainty High
uncertainty
relative to
full sample Long Short
Long-
Short Long Short
Long-
Short
(1) (2) (3) (4) (5) (6) (7)
1 1.09 -0.35 1.44* -0.46 0.30 -0.76 1.18**
2 1.36*** -0.12 1.48** 0.13 0.33 -0.20 1.17*
3 1.87*** -0.58 2.46*** 0.20 0.19 0.01 1.48**
4 1.06*** 0.04 1.02 -0.27 0.35 -0.62 0.81
5 0.90*** 0.18 0.72 -0.05 -0.35 0.30 0.43
6 1.19*** -0.40 1.58*** -0.03 -0.20 0.17 1.19***
7 0.97*** -0.35 1.42*** -0.16 0.07 -0.23 1.10***
8 0.97*** -0.48* 1.45*** -0.18 -0.22 0.04 1.18***
9 0.50*** -0.56** 1.06*** -0.14 -0.22 0.09 0.80**
10 0.42*** -0.44** 0.85*** -0.08 -0.33 0.25 0.59**
11 0.41*** -0.44** 0.85*** 0.03 -0.4** 0.37* 0.32*
52
Table 8: Analysts’ outputs at the aggregate level The dependent variables are the monthly value-weighted market returns (Panel A) and the returns on the VIX (Panel B)
for the calendar month following the month with the earnings announcement. The main explanatory variables in Panel A
are based on the recommendation revisions during the earnings announcement month, defined in Table 5. The consensus,
HQ, and LQ mean recommendation revision variables are the sums of all, HQ, and LQ analysts’ revisions, respectively,
divided by the number of revisions for all firms with an earnings announcement in that calendar month. The HQ-
consensus mean recommendation revision is the HQ mean recommendation revision minus the consensus mean
recommendation revision if the HQ mean recommendation revision is either positive and greater or negative and smaller
than the consensus recommendation revision; otherwise, this variable is zero. In Panel B, forecast dispersion is a value-
weighted average of firm-level analysts’ forecast error dispersions prior to the earnings announcement, defined in Table
6, with the firms’ market capitalizations as the weights. Analysts are ranked as HQ or LQ based on this year’s (Panel A)
and the previous year’s (Panel B) absolute forecast errors, defined in Table 1. The dispersion ratio of HQ (LQ) analysts
is the value-weighted average of firms’ ratios of HQ (LQ) analysts’ dispersion to Consensus dispersion. The control
variables are the monthly earnings-to-price ratio, dividend-to-price ratio, term spread, default spread, one-month T-bill
rate, 30-year Treasury yield, rate of inflation, and change in economic policy uncertainty (EPU), defined in Section 2.
The models use Newey-West standard errors with three lags in Panel A and Huber-White robust standard errors in Panel
B. The last row in Panel A reports the market model alphas obtained as follows. In Panel A, the alphas are from regressing
a long-short market-holding strategy return on the market return, where the market-holding strategy is the following. In
columns (5)-(7), long (short) the next month’s market return if the consensus, HQ, or LQ analysts’ mean recommendation
revision of that column is greater (smaller) than the median of the respective mean recommendation revision over the
previous 24 months; in column (8), long (short) the next month’s market return if the HQ-consensus mean
recommendation revision is positive (negative), and have a zero position in the market index if the HQ-consensus is zero.
The last line in Panel B provides the alphas from regressing the VIX strategy return on the market return, where the VIX
strategy return is a long or short VIX return, depending on whether the column’s dispersion or dispersion ratio last month
is greater or smaller than the median of the dispersion or dispersion ratio over the previous 24 months, respectively. t-
statistics are provided in parentheses. *, **, and *** indicate significance at the 10%, 5%, and 1% level, respectively.
53
Panel A: Predicting market returns
(1) (2) (3) (4) (5) (6) (7) (8)
Consensus mean 0.016* 0.015*
recommendation revision (1.85) (1.87)
HQ analysts’ mean 0.017** 0.017**
recommendation revision (2.25) (2.44)
LQ analysts’ mean 0.005 0.004
recommendation revision (0.91) (0.70)
HQ-consensus mean 0.027* 0.028*
recommendation revision (1.89) (1.89)
Lagged dependent 0.073 0.073 0.070 0.068 0.050 0.055 0.047 0.050
variable (0.93) (0.96) (0.90) (0.91) (0.73) (0.80) (0.68) (0.75)
Earnings-to-price ratio -0.148 -0.120 -0.175 -0.140
(-0.51) (-0.41) (-0.60) (-0.47)
Dividend-to-price ratio 1.913** 1.869** 1.968** 1.936**
(2.45) (2.40) (2.48) (2.45)
Term spread 0.679** 0.665* 0.647* 0.614*
(1.99) (1.94) (1.93) (1.81)
Default spread -1.733 -1.583 -1.857* -1.678
(-1.65) (-1.51) (-1.75) (-1.58)
One-month T-bill yield 3.308 3.136 3.323 3.025
(1.34) (1.24) (1.36) (1.18)
Long-term T-bond yield 0.091 0.100 0.090 0.105
(1.28) (1.44) (1.26) (1.52)
Inflation 0.470 0.437 0.544 0.499
(0.64) (0.61) (0.74) (0.70)
Intercept 0.009*** 0.010*** 0.008*** 0.008*** 0.028 0.026 0.026 0.022
(3.44) (3.59) (2.78) (2.93) (1.08) (1.02) (1.03) (0.90)
Observations 265 265 265 265 265 265 265 265
Adj. R-squared 0.09 0.018 -0.001 0.011 0.028 0.037 0.018 0.031
Monthly alpha (%) 0.42
(1.44)
0.49*
(1.65)
0.31
(1.06)
0.57***
(2.60)
54
Panel B: Predicting changes in market volatility
(1) (2) (3) (4) (5)
Consensus dispersion 2.104
(0.35)
HQ analysts’ dispersion 5.345
(0.64)
LQ analysts’ dispersion 2.625
(0.39)
Dispersion ratio of HQ analysts 0.107*
(1.91)
Dispersion ratio of LQ analysts -0.010
(-0.28)
Lagged dependent variable -0.046 -0.045 -0.047 -0.040 -0.047
(-0.36) (-0.36) (-0.36) (-0.30) (-0.36)
EPU change 0.060 0.062 0.061 0.068 0.059
(1.00) (1.04) (1.02) (1.16) (0.98)
Value-weighted market return 0.576 0.592 0.591 0.518 0.545
(1.40) (1.44) (1.43) (1.24) (1.33)
Earnings-to-price ratio -1.109 -1.075 -1.078 -1.560 -1.195
(-0.80) (-0.77) (-0.77) (-1.07) (-0.86)
Dividend-to-price ratio 1.244 0.865 1.092 3.567 1.919
(0.31) (0.22) (0.26) (0.88) (0.49)
Term spread 1.130 1.020 1.044 1.828 1.363
(0.57) (0.51) (0.51) (0.87) (0.67)
Default spread -4.540 -4.347 -4.523 -6.245 -5.034
(-0.85) (-0.82) (-0.85) (-1.14) (-0.94)
One-month T-bill yield 7.197 6.844 6.563 8.362 7.834
(0.52) (0.49) (0.47) (0.60) (0.55)
Long-term T-bond yield -0.150 -0.176 -0.134 -0.246 -0.155
(-0.37) (-0.43) (-0.33) (-0.61) (-0.38)
Inflation -2.114 -2.338 -2.194 -1.974 -1.832
(-0.53) (-0.58) (-0.54) (-0.50) (-0.46)
Intercept 0.127 0.120 0.123 0.075 0.150
(0.77) (0.73) (0.74) (0.47) (0.87)
Observations 197 197 197 197 197
Adj. R-squared (%) 0.038 0.041 0.038 0.056 0.037
Obs. in Long portfolio 115 124 74 141 64
Obs. in Short portfolio 97 88 138 71 148
Monthly alpha (%) 1.85 2.04 -0.57 3.32*** -0.56
(1.34) (1.60) (-0.40) (2.68) (-0.41)